fix: vectorize 3-bit and 6-bit bitstream packing (~1900× speedup)

[pbertsch]

Summary

• Replace Python for loops in _pack_bitstream / _unpack_bitstream with NumPy broadcasting
• Same public API, same wire format — zero behavioral change
• _pack_2bit and _pack_4bit were already vectorized; this brings 3-bit and 6-bit to the same level

Root cause

LCM(3,8) = 24 means each 3-bit block spans 8 indices × 3 bytes. The old implementation looped over every block in Python — O(n / block_size) interpreter ticks. For a typical LLaMA-3 8B KV cache (16M elements), that's 2M Python iterations per serialize call.

Fix

# Pack (n_full blocks at once)
blocks = flat[:n_full * block_indices].reshape(n_full, block_indices).astype(np.uint32)
acc = (blocks << idx_shifts).sum(axis=1, dtype=np.uint32)
bytes_2d = ((acc[:, None] >> byte_shifts) & 0xFF).astype(np.uint8)
out[:n_full * block_bytes] = bytes_2d.ravel()

Tail elements (< 8 for 3-bit) still use a tiny Python loop — they're negligible.

Benchmark (CPU, 131 072 elements, bits=3)

	Time
Before	~444 ms (from serde bench)
After	0.23 ms
Speedup	~1 900×

Test plan

• All 139 packing tests pass (pytest tests/test_packing.py)
• Full 787-test suite passes
• Manual roundtrip verified for bits ∈ {3, 6} at n ∈ {131 072, 1 048 576}

🤖 Generated with Claude Code

[pbertsch]

CI failures are pre-existing and unrelated to this PR:

• lint: F541 in benchmarks/benchmark_v05.py (untouched by this PR) — pre-existing since April
• macOS Metal GPU hang: flaky MLX test in test_compressed_cache.py — same failure on main for all recent commits

Linux 3.10/3.11/3.12 all pass. The packing change itself is a single-file edit with 787/787 tests green.