[TRAIN-6] flash-linear-attention and causal-conv1d missing from container and docs for Qwen3.5

## Problem

`flash-linear-attention` and `causal-conv1d` are required for full Qwen3.5 performance on the DTensor backend. Without these packages, DTensor silently falls back to the default PyTorch implementation and trains approximately 2x slower. Neither the NeMo-RL container nor the documentation mentions this requirement.

## Suggested fix

- Add `flash-linear-attention` and `causal-conv1d` to the NeMo-RL container.
- Document these packages as required dependencies for Qwen3.5 training on DTensor.