refactor(fspy-shm): split platform implementations

[wan9chi]

Motivation

fspy_shm's Unix and Windows implementations are diverging as the Unix attach path becomes usable before libc initialization. Keeping both behind platform conditionals in one implementation makes each incremental Unix change rewrite or conditionalize Windows code. This PR separates the implementations into unix.rs and windows.rs while retaining one public API and one shared cross-platform behavioral test suite.

[github-actions]

fspy benchmark

linux

dynamic/launch             change  +0.73%  [ -3.58% ..  +4.95%]  overhead   +52.19%
dynamic/access             change  -0.20%  [ -6.11% ..  +2.17%]  overhead    +5.89%
dynamic/access-relative    change  -0.08%  [ -1.81% ..  +2.30%]  overhead   +52.07%
static/launch              change  +0.59%  [ -3.04% ..  +5.32%]  overhead  +150.42%
static/access              change  -0.36%  [ -1.55% ..  +1.00%]  overhead  +736.93%
static/access-relative     change  +0.51%  [ -1.50% ..  +4.09%]  overhead +1177.35%

macos

dynamic/launch             change  -0.35%  [ -3.30% ..  +2.72%]  overhead  +214.78%
dynamic/access             change  +0.29%  [ -1.17% ..  +2.76%]  overhead    +6.62%
dynamic/access-relative    change  +0.44%  [ -2.10% ..  +2.88%]  overhead  +269.43%

windows

dynamic/launch             change  +0.14%  [ -1.47% ..  +2.02%]  overhead   +26.99%
dynamic/access             change  +0.19%  [ -0.74% ..  +0.99%]  overhead    +1.13%
dynamic/access-relative    change  +0.18%  [ -0.91% ..  +1.11%]  overhead    +1.28%