Implements "extended registers"

Extended registers provide access to a separate namespace of registers
backed by custom lua callback functions.

A new special register ("~) is introduced, which prompts the user for
the name of an extregistr to yank or paste from.

This enables file-backed or cloud-synced or other arbitrary register
behavior.

One can now use:

`"~file_extreg<cr>yy`

to yank into the "file_extreg" register.

AI-assisted: Qwen3-Coder

fix(lua): regression with `print()` in headless mode when redirecting

This is a bunch of hacks so we can override msg_puts_printf specifically
while still allowing redirecting and "normal" message flow to happen,
as this still has observable side-effects (even with no screen attached,
screenstring() is still valid..)

fixes #42218

Problem:
`vim._with({ buf = 0, win = 0 }, f)` fails an assert in ctx_switch().
vim._with() allows `buf` with `win` when `win` shows `buf`, but
nlua_with() passes both, and ctx_switch() takes only one target.

Solution:
Ignore `buf` if `win` is given; the window implies its buffer.

consider the following test_script.lua :

    print 'hello\n' -- TWO new lines
    print 'this is \027[31mRED FOR AN AMAZING REASON\027[0m'
    print 'foo'
    print '' -- newline despite empty!
    print 'bar'

    -- a bit later:
    os.execute 'sleep 2' -- CPU goes BRR, cursor in column zero
    print 'more bar'

before this PR, `nvim -l test_script.lua` had some weird behaviors
compared to `lua test_script.lua` or the now removed `nvim -ll
test_script.lua`.

- one extra '\n' at the end of a print is ignored, one would need
  TWO '\n' for a single extra '\n'

- an empty print'' is supposed to print an empty line

- print'\027[31mRED' is being escaped so it does not work

- the cursor gets stuck at the end of a message, instead of after
  emitting the new line, this looks like the CPU lag has something
  to do with the _last_ printed message which is likely not the case.

I consider these behaviors to be unintentional (a consequence of
stragglig through the message.c meta-machine, which is optimized
for message handling in _an interative text editor_) and getting
closer to plain lua behavior would be better.

An alternative would be to restore the builtin print(), but our
implemenation still has some benefits, like try
`print(setmetatable({}, {__tostring=error})` in plain lua..

Problem: Command inputs and parsed modifiers have different optional
fields, and completion window IDs are annotated as floating-point
values.

Solution: Separate modifier input and return types, correct integer IDs,
and update the metadata generators and their output.

AI-assisted

CI is currently using clang-tidy 21 but prepare for the future