chore(wandb): report GPU system metrics (SM Active) on secondaries

[Rockdu]

What

• Bump wandb from 0.23.1 → 0.28.1.
• Stop disabling system-stats collection on wandb secondaries (drop
  x_disable_stats=True in init_wandb_secondary).

Together these make the W&B System panel report the full per-GPU metric
set — including the NVML GPM family (GPU SM Active %, tensor/FP-pipe
activity, DRAM bandwidth, PCIe/NVLink throughput) — bringing us to parity with
miles core.

Why

Our System panel showed far fewer metrics than miles core (e.g. no
GPU SM Active (%)). Two independent causes, both fixed here:

1. Version. SM Active and the rest of the NVML GPM profiling metrics for
   Hopper+ GPUs are collected by wandb-core without a DCGM daemon, but the
   feature only landed in wandb v0.26.0 (2026-04-13, feat: add NVML GPM profiling metrics for Hopper+ GPUs to system monitor wandb/wandb#11622).
   We were pinned to 0.23.1 (2025-12-03), which predates it by ~4 months, so
   the collection path simply did not exist. miles core leaves wandb
   unpinned and floats to a version that has it.

2. Config. GPM metrics are sampled on the process that owns the GPU. In
   our Ray setup the wandb primary is the CPU-only driver
   (train_diffusion.py), which has no GPU; every GPU worker inits as a
   secondary, and those secondaries were forcing x_disable_stats=True —
   so no process was ever sampling GPU stats. miles core never sets this flag.

Only two single processes init as secondary — the training actor at global
rank 0 and the rollout manager — so re-enabling stats does not create
duplicate per-node sampling.

Note: SM Active requires Hopper+ (H100/H200 and newer) hardware, since
NVML GPM is only available there.

Files

• requirements.txt — bump wandb==0.23.1 → wandb==0.28.1.
• miles/utils/wandb_utils.py — drop x_disable_stats=True from the secondary
  settings; add a comment explaining why stats stay on.

Checklist

• pre-commit run --files ... passes on the changed files (yaml/ruff/autoflake/isort/format all green)
• Added/updated tests for new behaviour — no tests added; this repo has no wandb test module and the change is init-time W&B Settings wiring that can't be meaningfully unit-tested without a live wandb backend + GPU
• pytest -x is green — not run (no relevant tests; change is not unit-testable offline)
• If launch flags changed, python3 train.py --help still parses — n/a, no launch flags changed
• If a public flag was added, it appears in the CLI reference docs — n/a, no public flag added
• If an example was added, it has a real walkthrough — n/a, no example added
• Full verification requires an online run on H100/H200 to confirm GPU SM Active (%) and the rest of the GPM family appear in the System panel — not run (needs GPU + W&B backend)

[Rockdu]

Superseded by a draft PR against the upstream radixark/miles_diffusion repo.