Restructure Slurm training recipes to align with standard naming convention

[weikuo0506]

This PR restructures the Slurm-based training recipes in the training/ directory to comply with the new standardized folder and file naming convention. This follows the previous work done for GKE recipes.

Key Changes:

• 7-Level Directory Structure: Renamed Slurm directories to follow the standard: training/{accelerator}/{model}/{framework_runtime}/{version}/{scale_params}/recipe.
• Framework Normalization: Standardized framework names to *-slurm (e.g., megatron-bridge-slurm, nemo-slurm).
• Scale Parameters Alignment:
  • Used continuous precision naming (e.g., fp8cs, fp8mx) to match GKE.
  • Omitted seq{len} for all NeMo recipes and cases where it was not explicitly specified in the configuration files (e.g., Llama 3.1 405B Slurm recipes).
• Parameter Corrections based on File Content:
  • Corrected Global Batch Size (GBS) for qwen3_30b_a3b from 1024 to 512 based on the actual launch_script.sh.
  • Corrected total GPU count for wan_14b from 64 to 32 (8 nodes * 4 GPUs/node) based on wan_14b_benchmark.sh.
• Reference Updates: Updated internal path references in relevant scripts, readmes, and SLURM files.
• Cleanup: Removed empty parent directories left after moving files.

Scope:

This PR focuses strictly on Slurm recipes across a4, a4x, and a3ultra accelerators. GKE recipes in a3ultra and a3mega are left untouched for now.