Part 1 decode: marker-bit dequantization and block_states removal

[osamu620]

Context

j2k_dequant() reads the per-sample decoded-bit-plane index p from block_states (block_dequant.cpp:61, 91):

```cpp
N_b = 30 - (state >> 3) + 1;
```

This is the only remaining dependency on block_states once #315 is done. If we can get p from somewhere else, the block_states buffer can be deleted entirely — saving one allocation/zero per codeblock and its ~4 KB of memory traffic in the hot path.

Book recipe (§17.1.3)

Instead of storing p in a separate byte-per-sample array, embed it in sample_buf itself via a marker bit:

• When a sample first becomes significant (σ transitions 0 → 1), set a marker bit in sample_buf[j2 + j1*sstride] at position pLSB - 1 (one below the current bit-plane's LSB). This happens in sigprop's samples[...] |= 1 << p path and cleanup's equivalent.
• Dequant recovers p by counting trailing zeros of the sample (after masking off the sign bit), which identifies the lowest set bit — the marker.

Scope

• block_decoding.cpp: at each σ=1 transition, also OR the marker bit (1 << (p - 1) is fine in the natural flow since p >= 1 at decode time).
• block_dequant.cpp + NEON/AVX2/AVX512 variants: replace the block_states byte load with a trailing-zero-count on the decoded sample.
• coding_units.cpp + subband_row_buf.cpp: delete block_states / blkstate_stride / all allocation and zeroing for the Part 1 decode path. HT keeps its own block_states (different bit semantics, out of scope).
• j2k_codeblock: delete block_states / blkstate_stride fields (Part 1 only — HT uses a separate allocation path after Part 1 decode: full consumer port of σ/σ̄/π/χ̂ onto packed stripe-column word #315 migrates them).

Why this is not shipped yet

Without #315 in place, block_states still carries σ/σ̄/π/χ̂ for several consumers, so deleting it isn't possible. D alone on top of the current branch would save only the single block_states byte load in dequant — dequant is already SIMD'd, so the ceiling is ~2–5 % and the cost in touching all dequant kernels isn't justified.

D becomes attractive once #315 is merged: at that point block_states holds only p-index, and D lets us delete the whole buffer.

Dependency graph

• Blocked by: Part 1 decode: full consumer port of σ/σ̄/π/χ̂ onto packed stripe-column word #315 (full consumer port).
• Enables: deleting block_states for Part 1 entirely — one aligned_mem_alloc + one memset per precinct per DWT level gone, plus the corresponding memory footprint (up to a few MB for large tiles).
• Not applicable to: Part 15 HT (different block_states semantics, own allocation path).

Acceptance

• Conformance: 582/582.
• Perf: end-to-end Part 1 decode ≥ parity with the post-Part 1 decode: full consumer port of σ/σ̄/π/χ̂ onto packed stripe-column word #315 baseline; any dequant kernel regression must be offset by the removal of block_states allocation / zeroing / loads.
• Verify the marker-bit correctness against p0_07.j2k / p0_08.j2k (many bit-planes, worst case for marker-bit recovery).