fix(ext/ffi): trampoline for fast calls

[littledivy]

Pointed out by @aapoalas #15125 introduced a bug in FFI calls. The first parameter of a fast function must be a V8Value receiver. This patch implements a fix by patching symbol calls to ignore the receiver.

[aapoalas]

Nice! I like seeing tests with this.

[aapoalas]

Oh yeah, regarding the u8 / u16 / i8 / i16 support, @phosra had a good note about return values:
#15122 (comment)

[aapoalas]

If the first parameter is now the receiver, does this mean that C functions with empty parameters would now supported?

0-ary C functions were already supported. eg. Here're a few lines from benchmarks I ran on a PR of mine which changes nothing related to normal FFI calls:

benchmark                              time (avg)             (min … max)       p75       p99      p995
------------------------------------------------------------------------- -----------------------------
nop()                                3.56 ns/iter    (3.45 ns … 19.23 ns)    3.5 ns   4.37 ns   4.75 ns
add_u32()                            7.91 ns/iter    (7.17 ns … 18.92 ns)   7.64 ns  13.32 ns  15.29 ns
nop_u8()                             7.88 ns/iter     (7.4 ns … 34.05 ns)    7.4 ns  12.53 ns  16.59 ns

This just changes how those calls are called: Currently they're given an extra parameter that they do not expect but luckily also do not care about.

[ry]

It would be very nice to have ext/ffi/README updated with an summary of how this works.

[ry]

Please add an explanation of the process to ext/ffi/README

[ry]

Why the third_party changes if windows is not supported?

[ry]

LGTM - nice work Divy

TODO: remove tcc and directly write asm instead.