[zlib] Implement deflate wide writes

All the way back to 2017, we implemented wide reads/writes in *inflate*
on 64-bit Arm and x86 CPUs to speed up data decompression.

It turns out that the same idea can be exploited to speed up deflate
encoding on 64-bit CPUs.

This patch changes put_short & send_bits to write more data with less
instructions while performing compression by widening the bitstream
buffer from 16 to 64 bits. On x86-64 it uses a single 'MOV WORD' instead
of a pair of 'MOV BYTE' (similarly on aarch64: a single 'strh' instead
of a 'str' plus 'strb' pair).

The optimization is little endian specific, so the use of naming of the
macro (i.e. DEFLATE_CHUNK_WRITE_64LE). In theory it could be ported to
32bit, but we no longer add new optimizations to the 32bit code path,
only try to not break it.

The improvement will be more evident on lower compression levels (1 to
5) than on higher levels, as then more cycles are spent on searching
matches than writing data.

For Xeon 6th gen (Granite Rapids: GNR) the average compression speed
improvement ranges from +4.3% (default) to +8.5% (level 1). For Zen 5
the numbers are similar and so is for aarch64. For RISC-V it was a
lesser gain (+2.8% default compression up to +4.5% for level 1).

Original patch by Noel Gordon.

Bug: 419596999
Change-Id: Idb9a6f39c1ef4cc6f22bfb6fbaeec4d87ea17dc5
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/6580406
Commit-Queue: Adenilson Cavalcanti <cavalcantii@chromium.org>
Reviewed-by: Hans Wennborg <hans@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1468234}