[TheMeinerLP]

Problem

A train job with raw_hier_anomaly on a small snapshot was OOMKilled in an 8Gi pod after
the 0.5.0 default of 2048 packets per window. The packet transformer attends over all packets,
so activation memory grows with max_packets squared.

Measurements (P = 2048, default model, peak RSS growth)

Stage	Before	After
training step / LR range test, batch 16	~12.4 GiB (6.5 GiB at batch 8, measured)	3.4 GiB (micro-batch 4)
scoring, score_batch 16	~4.2 GiB (1.05 GiB at 4)	3.4 GiB (micro-batch 13)
validation pass over a split	~0.26 GiB per window, all at once	capped like scoring
ONNX parity run, 64 windows	16.8 GiB	3.8 GiB (13 rows per run)

Change

• corvus_ml/membudget.py: documented per-window activation estimate (attention + activations),
  budget (memory_budget_mb, CORVUS_ML_MEMORY_BUDGET_MB, else half of the cgroup memory.max,
  else RAM) and the micro-batch cap. A reduced micro-batch is logged.
• fit, find_lr and pretrain accumulate gradients over micro-batches, so batch_size stays
  the optimiser batch. The LR range test uses the candidate batch and the same micro-batch.
• Validation, scoring, the kNN memory bank (now under no_grad), the ONNX parity run and the
  benchmark batch are capped too.
• flatten_windows and pretrain no longer copy the window tensor when not needed.
• ml/scripts/measure_raw_memory.py reproduces the numbers.
• Docs: memory behaviour and rule of thumb in raw-byte-tokenizer.md, parameter in ml-jobs.md.

The masked-token loss is normalised per micro-batch, so accumulation is exact for mean losses
and approximate for it.