Revert torch version to 2.5.1

[hamadfadi]

Restores torch==2.5.1+cu121 and its corresponding wheel hashes in user_examples/llm_fine_tuning_cloud_batch/requirements.txt.

### Motivation & Root Cause Analysis

Attempting to upgrade to `torch==2.13.0` introduced critical runtime and build failures on the CUDA 12 infrastructure stack:

1. **C-Symbol Linkage Mismatch (`ncclCommResume`)**:
   - The platform base container (`infrastructure/requirements.txt`) installs CUDA 12 drivers and `nvidia-nccl-cu12` (v2.21.5).
   - `torch 2.13.0` requires CUDA 13 and `nvidia-nccl-cu13` (NCCL 2.29+), which calls the new `ncclCommResume` C-symbol.
   - At runtime, Python's dynamic loader loads `libnccl.so.2` from `nvidia-nccl-cu12` first. When `libtorch_cuda.so` attempts GPU initialization, it fails with:
     ```text
     /usr/local/lib/python3.12/site-packages/torch/lib/libtorch_cuda.so: undefined symbol: ncclCommResume
     ```
   - This causes GPU training loops to fail and silently fall back to surrogate mode.

2. **PyPI `--require-hashes` Version Lock Conflict**:
   - PyPI's official CUDA 13 wheels have mutually exclusive internal pins (`cuda-toolkit 13.0.3` pins `nvidia-cublas==13.1.1.3.*` while `nvidia-cudnn-cu13` pins `nvidia-cublas==13.6.1.10`). Under strict `--require-hashes` mode, `pip` cannot resolve the dependency graph.

### Solution
Restoring `torch==2.5.1+cu121` aligns PyTorch with the CUDA 12 base libraries (`nvidia-nccl-cu12`), eliminating C-symbol mismatches and ensuring real GPU fine-tuning executes properly on Cloud Batch worker VMs.