gh-156745: Inline CJK replacement writer call

[brittanyrey]

Call _PyUnicodeWriter_WriteCharInline() instead of the python lib version of the method.
Removes the overhead of the boundary call between the .so and the exported function.

# decode dense invalid bytes with errors='replace'
bad = b"\x81\x40" + b"\xff"*4000
bad.decode('shift_jis', 'replace')

Benchmark	base	patched	delta
replace, all-invalid (shift_jis)	34.06 µs	27.08 µs	−20.5%
replace, all-invalid (gbk)	36.55 µs	28.20 µs	−22.8%
replace, mixed valid/invalid	18.95 µs	16.85 µs	−11.1%
valid input, no errors (control)	18.50 µs	18.39 µs	−0.6% (noise)
ignore, all-invalid (control)	24.17 µs	24.16 µs	0.0% (noise)

• Issue: Performance: Optimize cross boundary call for codec.decode() error replace calls #156745

[corona10]

lgtm!