perf: major encode/decode optimization via const copy_from_slice and batch zero fill

Encode: rewrite to two-pass (scan + copy_from_slice) leveraging Rust 1.93's
const fn copy_from_slice for memcpy at runtime while keeping const fn encode.
Sub-slice scan pattern helps LLVM eliminate bounds checks. Added zero byte
fast path for all-zeros workloads. Now leads corncobs on nonzero/mixed encode.

Decode: batch consecutive 0x01 codes (all-zeros data) and fill zeros via
memset instead of per-byte loop. Decode zeros throughput jumps from ~1.2 GB/s
to ~4.6 GB/s, now 3x faster than corncobs on that workload.

Bump MSRV 1.83 → 1.93 for const copy_from_slice.
Update dev-deps: cobs 0.3→0.5, criterion 0.5→0.8, rand 0.8→0.10.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

benches/bench.rs

@@ -1,4 +1,6 @@
-use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
+use std::hint::black_box;
+
+use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
 
 // ── Test payloads ───────────────────────────────────────────────────
 
@@ -51,9 +53,11 @@ fn cobs_encode(data: &[u8]) -> Vec<u8> {
 }
 
 fn cobs_decode(encoded: &[u8], out: &mut [u8]) -> usize {
-    // unwrap_or(0): the cobs crate returns Err(EmptyFrame) for empty input,
-    // which is valid COBS. We treat it as 0 decoded bytes for benchmarking.
-    cobs::decode(encoded, out).unwrap_or(0)
+    // The cobs crate returns DecodeReport on success or Err(EmptyFrame)
+    // for empty input. We extract parsed_size, treating errors as 0.
+    cobs::decode(encoded, out)
+        .map(|report| report.parsed_size())
+        .unwrap_or(0)
 }
 
 // ── Benchmarks ──────────────────────────────────────────────────────