Problem:
`Completor:accept()` applies the item against the range the server
answered with. Characters typed after the request but before accepting
lie inside neither that range nor the replacement, so they are left
behind by the insertion: typing `a` after `foob` and accepting `foobar`
within the 200 ms debounce yields `foobara`.

Solution:
Grow the item's range over the characters typed since the request, so
that `accept()` replaces them. Only while the candidate still matches,
decided by the longest common prefix `show()` already computes, so input
that contradicts it is left alone. Snippet items are unaffected, since
`accept()` ignores the range for them.

Problem:
`Completor:accept()` hands the item's range straight to
`nvim_buf_set_text()`. That range is resolved when the response arrives,
so text deleted before accepting (backspacing within the 200 ms
debounce) leaves the range end past what is left of the line, and the
accept aborts with `Invalid 'end_col': out of range`.

18d6436 fixed the same staleness in `Completor:show()`, which drops an
item once its range *start* falls outside the buffer. That guard does
not cover the range end, and `accept()` was never given one, so this is
a hole left by that fix rather than a regression of it.

Solution:
Clamp the end of the range to the end of the line before writing. The
start needs no clamp: `show()` dropped the item one event loop tick
earlier if it had gone out of range.

Problem:
The magic globals `it`, `describe`, etc., are more trouble than they are
worth.

- Hooking into `after_each` requires `getfenv()` hacks.
- They confuse luals/emmylua, because the top-level `.luarc.json` isn't
  merged with `test/.luarc.json` (apparently a luals limitation?)
- They totally defeat discoverability because the user just has to
  "know" about the various magic symbols.

So they harm DX, which means they serve no purpose at all.

Solution:
- Expose the test API from `testutil`, so tests can call `t.it()`,
  `t.describe()`, etc., in the conventional way.
- Drop `getfenv()` hacks.
- Drop the `setfenv()` injection in `load_chunk`.
- Drop `test/_meta.lua`.

Problem:
Using nested `vim.Pos` objects to represent each `vim.Range` object
requires 3 tables for each `vim.Range`, which may be undesirable in
performance critical code. Using key-value tables performs worse than
using array-like tables (lists).

Solution:
Use array-like indices for the internal fields of both `vim.Pos` and
`vim.Range` objects. Use a metatable to allow users to access them like
if they were key-value tables.

---

Problem:
The `vim.Pos` conversion interface for `extmark` indexing does not take
into account the difference in how a position on top of a newline is
represented in `vim.Pos` and `extmark`.
- `vim.Pos`: for a newline at the end of row `n`, `row` takes the value
  `n + 1` and `col` takes the value `0`.
- `extmark`: for a newline at the end of for `n`, `row` takes the value
  `n` and `col` takes the value `#row_text`.

Solution:
Handle this in the `extmark` interface.

---

Problem:
Not all `to_xxx` interfaces have wrapping objects like `to_lsp`.

Solution:
Return unwrapped values in `to_xxx` interfaces where it makes sense.
Accept unwrapped values in "from" interfaces where it makes sense.

---

Problem:
`start` and `end` positions have different semantics, so they can't be
compared. `vim.Range` relies on comparing the `end` and `start` of two
ranges to decide which one is greater, which doesn't work as expected
because this of the different semantics.

For example, for the ranges:

    local a = {
      start = { row = 0, col = 22, },
      end_ = { row = 0, col = 24, },
    }
    local b = {
      start = { row = 0, col = 17, },
      end_ = { row = 0, col = 22, },
    }

in this code:

    local foo, bar = "foo",  "bar"
    --               |---||-|
    --                 b  a

The range `b` is smaller than the range `a`, but the current
implementation compares `b._end` (`col = 22`) and `a.start` (`col = 22`)
and concludes that, since `b.col` is not smaller than `a.col`, `b`
should be greater than `a`.

Solution:
- Use a `to_inclusive_pos` to normalize end positions inside of
  `vim.Range` whenever a comparison between a start and an end position
  is necessary.

* feat(lua): `Range:is_empty()` to check vim.range emptiness

* fix(lsp): don't overlay insertion-style inline completions

**Problem:** Some servers commonly respond with an empty inline
completion range which acts as a position where text should be inserted.
However, the inline completion module assumes that all responses with a
range are deletions + insertions that thus require an `overlay` display
style. This causes an incorrect preview, because the virtual text should
have the `inline` display style (to reflect that this is purely an
insertion).

**Solution:** Only use `overlay` for non-empty replacement ranges.

This also fixes the following warning in tests with ASAN or TSAN:

    -------- Running tests from test/functional/plugin/lsp/inline_completion_spec.lua
    RUN      T4604 vim.lsp.inline_completion enable() requests or abort when entered/left insert mode: 225.00 ms OK
    RUN      T4605 vim.lsp.inline_completion get() applies the current candidate: 212.00 ms OK
    nvim took 2013 milliseconds to exit after last test
    This indicates a likely problem with the test even if it passed!

    RUN      T4606 vim.lsp.inline_completion get() accepts on_accept callback: 212.00 ms OK
    RUN      T4607 vim.lsp.inline_completion select() selects the next candidate: 220.00 ms OK
    -------- 4 tests from test/functional/plugin/lsp/inline_completion_spec.lua (3437.00 ms total)

    -------- Running tests from test/functional/plugin/lsp/linked_editing_range_spec.lua
    nvim took 2011 milliseconds to exit after last test
    This indicates a likely problem with the test even if it passed!

The flakiness happens because get() uses vim.schedule(), and a following
key may be processed before the scheduled event. Use poke_eventloop() to
ensure that the scheduled event is processed.