fix(par2): skip PAR2 creation for files smaller than SIMD alignment

[javi11]

Summary

Fixes intermittent CI segfault in internal/par2:

INFO Starting par2 creation process executor=NativeExecutor
signal: segmentation fault (core dumped)
FAIL  github.com/javi11/postie/internal/par2  0.794s

Root cause

The par2go C backend may read up to stride-1 bytes past the input buffer when the slice size isn't a multiple of the SIMD stride (64 B on AVX-512). PR #189 fixed this for files ≥128 B by rounding the slice size down to a 128-byte boundary, but the else branch still passed sub-stride slices (e.g. 8 B for a 10-byte file) to the C library, causing segfaults on Linux x86_64 CI runners.

Crashes appeared "flaky" because the overrun only segfaults when the read crosses into an unmapped page — depends on heap layout, ASLR, and allocator state. The existing defer recover() in TestIntegration_NativeExecutor_HandlesVerySmallFiles_10Bytes cannot catch CGO segfaults (they bypass Go's panic mechanism). Locally on macOS the SIMD instruction path differs and the bug stays latent.

Change

In createPar2ForFile, when parBlockSize > file.Size and file.Size < 128, skip PAR2 creation with a warning instead of computing a sub-stride slice. Files that small yield no meaningful recovery data, and the upload pipeline already tolerates a nil par2 result (file upload runs in a parallel goroutine, poster.Post(ctx, nil, ...) is a no-op).

The existing 10-byte regression test was already written assuming "skip" is a valid outcome ("expect either a nil result (skipped with warning) or a graceful error, but never a panic").

Test plan

• go test -race -count=1 ./internal/par2/... passes locally
• TestIntegration_NativeExecutor_HandlesVerySmallFiles_10Bytes now takes the skip path
• TestIntegration_NativeExecutor_HandlesVerySmallFiles_512Bytes still creates PAR2 (≥128 B path unchanged)
• CI run on AVX-512 Linux: segfault gone