Codec benchmark harnesses: fail loudly on skips, empty corpora and AppleDouble files; time init; share zstd contexts per thread

Lessons from a binary-format encoder benchmark: (1) Exit nonzero on round-trip failure, unreadable input AND an empty corpus. A bench that prints FAIL but exits 0 silently passes in CI. (2) Folders copied from macOS often carry AppleDouble ._name files that match find -name "*.ext". Exclude them with ! -name "._*", and count unreadable files as failures, not silent skips (that is how we found them). (3) Valid zero-event files (header only) are edge cases: encode and compare them rather than skipping them. (4) Include construction and init in the timed region, or report both. Per-object preallocation can dominate with thousands of small encoders. (5) With many small zstd encoders (one per signal or stream), keep one ZSTD_CCtx plus scratch buffers per thread (thread_local workspace, grown to the largest block). Per-object contexts and preallocated buffers cost MBs each and add up to GBs at 10K objects. Output is unchanged because ZSTD_compress2 with the same parameters is deterministic regardless of context reuse.