Reduce default-flash RAM by reusing BF16 projection scratch

[Mikyx-1]

Default flash attention retains a full sequence-major projection history
after transposing it into the K/V buffers consumed by attention.
This change reuses BF16 projection scratch and reduces peak process RAM
while preserving the projection rounding and attention computation.

Fixes #1029

Implementation

• Reuse one layer of batch-sized BF16 projection scratch per query.
• Size it for the widest layer; grow it for larger batches or layer widths.
• Convert older-constructor caches when GemmaAttention first uses them.
• Preserve sequence capacity and the transposed K/V history.
• Copy the actual K/V buffers when cloning the new scratch layout.
• Skip unused compact-buffer allocations for the default-flash runtime path.
• Preserve normalization, RoPE positions, and both BF16 rounding boundaries.

Measured RAM effect

Peak RSS includes weights, KV caches, and inference working buffers.
These percentages describe the whole process, not model-weight compression.

Gemma 3 model	Prompt tokens	Before	After	Reduction
270M	32,736	1,778.59 MiB	1,205.35 MiB	32.2%
1B	16,352	2,369.89 MiB	1,957.65 MiB	17.4%
4B	4,064	5,754.25 MiB	5,216.35 MiB	9.3%

For 270M, the projection allocation falls from 578.00 to 4.25 MiB.
Combined projection plus K/V allocations fall from 1,154.00 to 580.25 MiB.
Dormant virtual allocations are not counted as physical RAM savings.

Measured speed effect

Model	Prefill seconds, before → after	Decode tokens/s, before → after
270M	39.4237 → 39.3970	37.4104 → 37.8506
1B	68.7325 → 68.5894	12.1040 → 11.9374
4B	74.3011 → 74.3155	2.9791 → 3.0151

Observed decode changes: +1.2% for 270M, -1.4% for 1B, +1.2% for 4B.
Prefill is essentially unchanged. The small measured speedups are tentative:
16-token decode intervals and few repeats do not establish statistical significance.

Measurement setup

• Baseline ffc1abc; measured fix caa2e13; Linux / Intel i5-12400F.
• Release AVX2/Haswell, six pinned threads, no oneDNN or spinning.
• 270M: capacity 32,768, batch 4,096, three stock-autotuning repetitions.
• 1B: capacity 16,384, batch 4,096; 4B: capacity 4,096, batch 1,024.
• 1B/4B: two repeats with fixed matmul candidates and forced continuation.
• 4B uses the existing padded checkpoint with --map 1 in both versions.
• Timings are medians; all runs are serial with a RAM/swap watchdog.
• Model snapshots show no swap; system swap grew 162.25 MiB in the 4B baseline.

Validation and scope

• Five focused cache/attention tests passed on AVX2.
• Coverage includes two queries, both constructor paths, heterogeneous widths,
  scratch growth, retained sequence capacity, cache copying, and T5 storage.
• Fixed-matmul verification: all 32 full-logit hashes match for each model.
• Each version also matches itself across two repetitions.
• That covers 8,388,608 logit values per model under identical continuation.
• T5, DeepSeek, and alternative tiled-backend inference storage are unchanged.
• No quantization, local-window rings, context truncation, or weight changes.
• The existing tiled test could not compile due to missing mock/span symbols.
• Only three implementation files and two test files are included.
• Benchmark artifacts, diagnostic harnesses, and reports are not committed.