Support final normalization of streamed Megatron gradients

[oran-goodman]

• Allow an optimizer request to scale accumulated gradients after distributed reduction and before clipping or Adam updates. This lets GSPO begin backward computation before the full batch denominator is known.
• Advertise streaming_accumulation_v1 only for non-colocated Megatron; default requests keep their existing behavior, and unsupported runtimes reject non-default scaling.

Testing: 14 CPU tests passed, including Adam parameters/moments and expert buffers; Ruff 0.11.9 and Black 24.10.0 passed. Qwen3.5-4B GPU qualification passed three updates with an interleaved adapter: gradient norms and outputs matched exactly, as did all 55.1M final main parameter entries and both Adam moment arrays with matched request packing. Different packing changes frozen-model numerics.

The Trajectory integration completed 20 steps on both configurations: streamed run 1050962 versus reference 1050942 reduced median step time 16.5% and weight-publication intervals 19.0%, with 9.8% higher accepted action-token throughput (initial step excluded). Both had zero stale-group rejections. This is one stochastic pair, not bitwise full-batch parity. The temporary runtime was shut down and checkpoints preserved.