fix(par2): SIMD-align slice size to prevent random AVX-512 segfaults

[javi11]

Summary

Fixes the random signal: segmentation fault (core dumped) in the Coverage workflow
on internal/par2 (e.g. run 26437870974 / job 77824805140).

Root cause

createPar2ForFile only rounded the slice size to a 4-byte boundary (per the PAR2 spec)
before handing it to par2go. The ParPar C kernel on Linux x86_64 with AVX-512 reads in
64-byte vector lanes, so a non-aligned slice size (e.g. articleSize = 10_000 →
SliceSize = 10_000, which is 156 * 64 + 16) causes overreads past the slice buffer.

The crash only fires when the over-read happens to cross an unmapped page. That depends
on the allocator's run-to-run layout — flaky on Linux CI, never reproducible locally on
macOS / arm64 (different SIMD stride, and the existing 128-byte clamp happened to mask it).

The source already documented the requirement at par2.go:430-444, but the SIMD-safe
alignment guard was only applied when parBlockSize > file.Size. For the failing path
(SliceSize = 10MiB > file.Size = 100KB → fallback → articleSize = 10_000) the guard
was skipped entirely.

Fix

• Extract per-file block-size policy into a testable computeFileBlockSize helper.
• Make 128-byte SIMD alignment unconditional for every slice size handed to par2go
  (both createPar2ForFile and computeSetBlockSize).
• Raise the "block too small" threshold from 4 to 128 and emit a clear "SIMD-safe"
  warning when skipping.
• Add TestComputeFileBlockSize_SIMDAligned covering the CI repro case plus several
  edge cases; asserts result % 128 == 0 so we can't regress silently.

Test plan

• go build ./internal/par2/...
• go test -race -count=3 ./internal/par2/... — passes
• TestComputeFileBlockSize_SIMDAligned covers the previously-crashing parameters
• CI Coverage workflow goes green on Linux x86_64