Every CJK decoder emits output through OUTCHAR(), which called the
exported _PyUnicodeWriter_WriteChar().  The codecs are shared extensions,
so that is a cross-module call for each output character, and the
function is nothing but the _PyUnicodeWriter_Prepare() macro plus a store
that the caller could have done itself.  OUTCHAR2, directly below it, was
already written inline for exactly this reason.

Use _PyUnicodeWriter_WriteCharInline() instead.  One macro, and it
applies to every codec in the package.

| benchmark               |    before |     after | change | spread |
|-------------------------|----------:|----------:|-------:|-------:|
| big5 decode 1.8MB       |   6.32 ms |   2.56 ms |   -59% |  0.3pp |
| euc_kr decode 1.8MB     |   6.94 ms |   3.18 ms |   -54% |  0.1pp |
| shift_jis decode 1.8MB  |   6.93 ms |   3.17 ms |   -54% |  0.4pp |
| euc_jp decode 1.8MB     |   6.32 ms |   3.18 ms |   -50% |  0.2pp |
| gb2312 decode 1.8MB     |   6.32 ms |   3.18 ms |   -50% |  0.3pp |
| euc_jp decode kana      |   3.26 ms |   2.05 ms |   -37% |  4.0pp |
| iso2022_jp decode       |   9.48 ms |   7.09 ms |   -25% |  8.3pp |

Through the StreamReader wrappers, where Python-level call overhead
dilutes it:

| benchmark                  |    before |     after | change |
|----------------------------|----------:|----------:|-------:|
| euc_jp read() [1 call]     |   6.33 ms |   3.18 ms |   -50% |
| euc_jp read(1024) xN       |   7.06 ms |   3.93 ms |   -44% |
| euc_jp readlines()         |   8.76 ms |   5.60 ms |   -36% |
| shift_jis readline() x20k  |  11.91 ms |   8.25 ms |   -30% |
| euc_jp readline() x20k     |  11.35 ms |   9.27 ms |   -18% |

Decoding every 1- and 2-byte sequence (65536 + 256 inputs) under each of
the 23 CJK codecs, plus 3-byte probes, hashes identically before and
after.

Measured with the two builds in separate worktrees.  These codecs live in
shared .so files, so copying the python executable within one build tree
does not isolate them: both copies load whichever module was built last,
and the change measures as an exact no-op.

…JhAM2.rst