docs(readme): document the create_trainer training providers

[sundargthb]

What

The README documents create_policy providers in depth (a dedicated Policy providers section with a provider table), but there is no equivalent for create_trainer, its training-side peer. The RL trainers added in #915 / #918 / #920 (create_trainer("ppo"), create_trainer("fast_sac")) only appear in the CHANGELOG.

This adds a Training providers section mirroring the Policy providers layout.

Contents

• create_trainer / TrainSpec usage (imitation post-tuning and from-scratch RL), matching the pattern in strands_robots/training/__init__.py.
• A provider table: lerobot_local, groot, cosmos3, mock (imitation) and ppo, fast_sac (RL).
• A note on the RL lane: BaseRLAlgo, the validate -> prepare -> train -> export lifecycle, VecSimEnv rollouts, and BaseRLAlgo.evaluate().

Verification

Every claim checked against source on main:

• provider names → register_trainer("ppo", ...) / register_trainer("fast_sac", ...) in training/__init__.py; imitation trainers in training/{lerobot,groot,cosmos3,mock}.py.
• create_trainer(provider, **kwargs) -> Trainer and the validate/prepare/train/export methods → training/base.py and training/factory.py.
• TrainSpec fields (dataset_root, base_model, output_dir, method, steps) → training/base.py; method values full | lora | ....
• VecSimEnv description → its class docstring and the feat(training): VecSimEnv - N independent SimEnv as one (N, D)-batched env #918 commit title.

Docs only, no code changes.

[yinsong1986]

Summary

Docs-only PR (+36/-0 in README.md) adding a Training providers section that mirrors the existing Policy providers layout. It documents create_trainer/TrainSpec for both imitation post-tuning and from-scratch RL, a provider table, and a note on the RL lane (BaseRLAlgo, the validate -> prepare -> train -> export lifecycle, VecSimEnv rollouts, BaseRLAlgo.evaluate(), and the torch-free-until-first-use property).

What's good

• Every claim checks out against source on main: create_trainer(provider, **kwargs) -> Trainer (training/factory.py:116); the four imitation providers' trainer blocks in registry/policies.json (mock/groot/lerobot_local/cosmos3); register_trainer("ppo"/"fast_sac", ...) in training/__init__.py:59,68; TrainSpec fields dataset_root/base_model/output_dir/method (full|lora|...)/steps in training/base.py; PpoTrainer(BaseRLAlgo) / FastSacTrainer(BaseRLAlgo); the VecSimEnv "N independent SimEnv as one batched env" wording matches its class docstring verbatim; train_policy wraps create_trainer (tools/train_policy.py:160,181); the groot embodiment-tag and cosmos3 multi-node HSDP notes are accurate.
• Scope discipline: pure documentation, no code touched.
• No emojis or non-ASCII in the added section, and no host paths — consistent with AGENTS.md string-hygiene and no-host-path conventions.