Fix script buffer leaks in coremini_load

[drebbe-intrepid]

Repeated coremini_load calls retained the entire script allocation, including on native failure. Own malloc-backed script buffers with a unique_ptr/free deleter in coremini_load and load_readbin so all returns and exception exits release storage. Check allocation failure, tuple length narrowing, and integer conversion errors while preserving byte conversion behavior and the native ABI.

Fixes #247

Validation:

• Rebuilt Windows x64 CPython 3.14.5 extension in an isolated worktree.
• Baseline: all four new file/tuple success/failure leak cases failed, retaining about 32 MiB over 32 calls.
• Fixed full suite: 34 passed, 4 skipped. Mock libraries validate payload bytes without hardware; tests cover native failures, conversion errors, empty inputs, and missing files.
• Four skips are specifically load_readbin: the public build does not expose this internal-header-only entry point. Its ownership fix was source-reviewed, not runtime-validated.
• Memory-growth assertions run on Windows; behavioral checks also support Linux/macOS mock libraries. Allocation failure and short-read fault injection were not run.
• Independent critic /root/critic247: "I explicitly agree with the final diff. No blocking findings." Independent focused tests: 4 passed, 4 skipped. Builder explicitly agrees with the same diff. Renewed review includes the Windows fixture correction; reviewed diff became commit 12bca98 (base 45167e0).
• Windows fixture targets Python bitness and exports undecorated i686 stdcall aliases; all child processes have 60s timeouts. Forced i686 fixture compilation and llvm-readobj verified PE I386 and both undecorated exports. No 32-bit Python runtime was exercised locally.
• git diff --check passed. CI pending; no hardware operations performed.

Parent coordinator retains CI acceptance, integration, hardware testing, and merge responsibility.