Default flash attention retains redundant BF16 KV projection history

## Summary

Default flash attention retains two BF16 representations of KV history.
The sequence-major projection buffer remains allocated and populated after
its keys and values have been transposed into the buffers attention reads.
This increases resident memory as context length grows.

## Affected path

- Backend: default `--attention_impl flash`.
- Confirmed with Gemma 3 270M, 1B, and 4B text inference.
- Measured baseline: `ffc1abc05abdf11d875d25d36c8859553ddf2641`.
- The retained-history path is also present on current `dev` (`b68def1`).
- T5 and DeepSeek use their legacy buffers differently.

## What happens

1. `ComputeQKV` writes BF16 projections into sequence-major `kv_cache`.
2. It applies normalization and positional encoding with BF16 rounding.
3. It transposes the results into `k_cache` and `v_cache`.
4. `FlashAttention` consumes those transposed buffers.
5. The original projection buffer still retains every layer and position.

References:
- [Projection and transpose path](https://github.com/google/gemma.cpp/blob/ffc1abc05abdf11d875d25d36c8859553ddf2641/gemma/attention.cc#L208-L322).
- [Full-sequence cache allocations](https://github.com/google/gemma.cpp/blob/ffc1abc05abdf11d875d25d36c8859553ddf2641/gemma/kv_cache.cc#L275-L297).

## Reproduction and observed cost

Run default-flash Gemma 3 270M with a 32,736-token text prompt,
sequence capacity 32,768, prefill batch 4,096, and 16 decode tokens.
Inspect the allocated cache extents and peak process RSS during inference.

The projection buffer occupies 578 MiB including row padding.
The transposed K/V buffers occupy another 576 MiB.
Measured peak process RSS is 1,778.59 MiB on this workload.
This is active retained data, not merely unused virtual address space.

Environment: Linux, Intel i5-12400F, six pinned threads,
Release AVX2/Haswell build, no oneDNN, and approximately 15.5 GiB RAM.

## Expected behavior

Projection intermediates should not retain a second complete KV history
after attention's persistent representation has been produced.
Required context, existing BF16 rounding, and model behavior must be preserved.

Lies zuerst die Zeilen 208-322 von gemma/attention.cc und die Zeilen 275-297 von gemma/kv_cache.cc, um die Projektion, Transponierung und Cache-Allokationen nachzuverfolgen. Reproduziere die default-flash-Gemma-3-Workload mit dem angegebenen Prompt und den angegebenen Sequenzeinstellungen und untersuche anschließend die Cache-Extents und den maximalen RSS. Als erledigt gilt die Aufgabe, wenn der redundante KV-Verlauf nicht beibehalten wird, während Kontextlänge, BF16-Rundung und Modellverhalten unverändert bleiben.

Vom Indexierungsmodell aus dem Issue-Text verfasst.