Merge pull request #2 from wudi/fix/issue-1-type-mismatch-errors

feat(errors): report struct.field path on type mismatch

jsonx v0.1.0 — first release

Drop-in `encoding/json` replacement with AVX-512BW (amd64) and NEON
(arm64) accelerated kernels. Decoder + encoder are both faster than
`bytedance/sonic` on every measured gate of the standard corpus.

Highlights (EPYC-Genoa AVX-512BW, 5s × count≥2 benchmarks):

  Decode interface{} — +10 % to +63 % over sonic on small / twitter /
  citm_catalog / canada / 1-5-10 MB deeply-formatted payloads.

  Decode struct — +50 % to +88 % over sonic on the same corpora
  (1 MB formatted: 1499 MB/s vs sonic's 797 MB/s).

  Encode interface{} — +11 % to +52 % over sonic across every
  corpus after the Schubfach formatter + amd64 digit-emission asm
  kernel work.

  ARM64 (Ampere Altra / Neoverse-N1): struct decode +26 %-+65 %
  over sonic; 1 MB deep-struct 789 MB/s vs sonic's 627 MB/s.

Architecture:
  - Type-specialized decoder plan cache (no reflect after warmup).
  - Slab allocators for interface{} scalar boxes + slice headers.
  - SWAR fast-paths on string scan, whitespace skip, number parse.
  - Single-compare whitespace kernel (VPCMPUB on amd64, CMHI on
    arm64) exploiting JSON's grammar: any structural-position byte
    ≤ 0x20 is whitespace or malformed.
  - Schubfach float formatter (pure Go, arch-agnostic); Eisel-Lemire
    decoder fed directly from the mantissa scan.
  - Encoder: one allocation per call; pre-quoted struct keys.

Full scorecard, experiment log, and reproducer benchmarks are in
RESULTS.md and README.md.

Requires Go 1.22+. No CGO.