docs: update homepage Core features section per the v0.1 feature list

[Shi-Dong]

Summary

• Rewrite the homepage Core features section (docs/index.md) to follow the Miles v0.1 feature list, splitting it into two groups:
  • Efficiency & stability — fully async RL, fast agentic rollout on SGLang + SGLang router, in-loop weight sync (<10 s for a Kimi-K2.6-class model, P2P RDMA fast path), unified low-precision training (MXFP8/NVFP4 stable RL recipe; FP8/INT4 QAT/BF16/FP16), TITO, Rollout Routing Replay (R3), LoRA & multi-LoRA, SGLang-engine fault tolerance, and day-0 model support.
  • Design, support & user experience — coding-agent sandboxes (Harbor / OpenEnv / NeMo-Gym with local CPU sandboxes or Daytona / E2B / self-hosted AgentENV providers), the plug-point customization surface, Megatron ↔ FSDP backend switch, recipe support (GRPO, PPO, SFT, on-policy distillation), verified hardware (GB300/GB200/B200/H200/H100/MI355X), and the Miles dashboard.
• Feature bullets link to their docs pages (Fully Async Rollout, P2P Weight Transfer, Low Precision, INT4 QAT, TITO/Agentic Chat Templates, R3, LoRA, Fault Tolerance, Harbor, OpenEnv, NeMo-Gym, Environments, Customization, FSDP backend, Platforms, Dashboard).
• Update the Advanced Features index intro (docs/advanced/index.md), which enumerates what the homepage Core-features section points at.

Notes:

• The feature list says "21 plug-points", but docs/user-guide/customization.md documents 26 --*-path hooks, so the page keeps the existing "twenty-plus" phrasing.
• The old "speculative rollout with online MTP-SFT" bullet is not part of the v0.1 list and is dropped from the homepage; the Speculative Decoding page itself is untouched and still linked from the Advanced tab and model pages.
• Sandbox providers in the sandboxes bullet are the ones with in-tree examples (examples/experimental/): Daytona, E2B, and AgentENV (the OpenEnv E2B leg also serves self-hosted AgentENV); there is no Modal integration in tree.

[guapisolo]

like fully-async scheduling and token info for each trajectory ?

[guapisolo]

Mentioned more algorithms here?