feat(training): add multi-GPU and mixed-precision support to OSMO LeRobot training workflow

[rezatnoMsirhC]

Summary

The training/il/workflows/osmo/lerobot-train.yaml OSMO workflow hardcodes gpu: 1 in its resources block and has no MIXED_PRECISION environment variable. This creates a feature parity gap with the AzureML equivalent (training/il/workflows/azureml/lerobot-train.yaml), which was updated in #778 to support configurable GPU count and mixed-precision mode.

A user submitting to an OSMO pool node with a multi-GPU SKU will silently get single-GPU training because gpu: 1 caps the container's device allocation regardless of the pool SKU. train.py also defaults MIXED_PRECISION to "no" when the variable is absent, so bf16/fp16 training is not accessible from OSMO submissions.

What needs to be done

Mirror the pattern already present in training/il/workflows/azureml/lerobot-train.yaml:

# resources block
resources:
  default:
    gpu: "{{ num_gpus }}"

# environment block
environment:
  MIXED_PRECISION: "{{ mixed_precision }}"

# default-values
default-values:
  num_gpus: 1
  mixed_precision: "no"

With num_gpus: 1 and mixed_precision: "no" as defaults, existing single-GPU submissions are unaffected.

Relevant files

• training/il/workflows/osmo/lerobot-train.yaml — change target
• training/il/workflows/azureml/lerobot-train.yaml — reference implementation
• training/il/scripts/lerobot/train.py — MIXED_PRECISION env var already read and validated here

Context

Identified during review of #778. The only change to the OSMO workflow in that PR was adding --no-cache-dir --no-deps to the uv pip install line (a lockfile correctness fix). Multi-GPU OSMO support was explicitly deferred as out of scope for that PR.