--deterministic-mode fails to ensure reproducibility in distributed training (Megatron-Core v0.14.0)

## Bug Description

The `--deterministic-mode` flag does not guarantee reproducible training results in Megatron-Core v0.14.0 when running multi-node distributed training. Despite identical configurations, seeds, and deterministic environment variables, loss values and model convergence differ across runs.

This issue persists even with comprehensive deterministic configurations including:
- `CUBLAS_WORKSPACE_CONFIG`
- `NVTE_ALLOW_NONDETERMINISTIC_ALGO=0`
- Disabled NVLS and Flash Attention
- `CUDA_DEVICE_MAX_CONNECTIONS=1`

---

## Environment

| Component | Version/Details |
|-----------|----------------|
| **Megatron-Core** | v0.14.0 |
| **Docker Image** | `nvcr.io/nvidia/pytorch:25.03-py3` |
| **GPU** | NVIDIA H200 |
| **Configuration** | 2 nodes × 8 GPUs, TP=8, PP=2 |
| **PyTorch** | [auto-detected from container] |
| **CUDA** | [auto-detected from container] |

---

## Reproduction Steps

### Configuration Details
- **Model**: GPT-style, 32 layers, 4096 hidden size
- **Parallelism**: TP=8, PP=2, Sequence Parallel enabled
- **Batch Size**: Micro-batch=32, Global-batch=64
- **Training**: 10 iterations with mock data
- **Precision**: BF16

### Launch Commands

**Node 0:**
```bash
export NCCL_ALGO=^NVLS
export NVTE_ALLOW_NONDETERMINISTIC_ALGO=0
export CUBLAS_WORKSPACE_CONFIG=:4096:8
export NCCL_NVLS_ENABLE=0
export NVTE_FUSED_ATTN=1
export NVTE_FLASH_ATTN=0
export CUDA_DEVICE_MAX_CONNECTIONS=1

torchrun \
--node_rank 0 \
--nnodes 2 \
--nproc_per_node gpu \
--master_addr 10.53.249.242 \
--master_port 23456 \
./pretrain_gpt.py \
--tensor-model-parallel-size 8 \
--pipeline-model-parallel-size 2 \
--num-layers 32 \
--hidden-size 4096 \
--num-attention-heads 32 \
--seq-length 2048 \
--max-position-embeddings 2048 \
--micro-batch-size 32 \
--global-batch-size 64 \
--no-overlap-p2p-communication \
--train-iters 10 \
--deterministic-mode \
--attention-backend fused \
--use-distributed-optimizer \
--sequence-parallel \
--log-interval 1 \
--log-throughput \
--no-one-logger \
--lr-decay-iters 35 \
--lr-warmup-iters 5 \
--distributed-backend nccl \
--lr 6.0e-5 \
--min-lr 6.0e-6 \
--lr-decay-style cosine \
--merge-file /sanitycheck/data/megatron/dataset/gpt2-merges.txt \
--vocab-file /sanitycheck/data/megatron/dataset/gpt2-vocab.json \
--clip-grad 1.0 \
--weight-decay 0.1 \
--adam-beta1 0.9 \
--adam-beta2 0.95 \
--num-workers 4 \
--init-method-std 0.006 \
--bf16 \
--mock-data \
--eval-iters 0 \
--rerun-mode disabled
```

**Node 1:**
```bash
# Same environment variables as Node 0
torchrun \
--node_rank 1 \
--nnodes 2 \
--nproc_per_node gpu \
--master_addr 10.53.249.242 \
--master_port 23456 \
./pretrain_gpt.py \
[... identical arguments as Node 0 ...]
```

---

## Expected Behavior

Running the same command multiple times with `--deterministic-mode` should produce:
- ✅ Identical loss values at each iteration
- ✅ Bit-exact model weights/gradients
- ✅ Identical convergence trajectory
- ✅ Reproducible checkpoints

---

## Actual Behavior

❌ Loss and gradient norm values are non-deterministic across runs:

**Example:**
```
Run 1 - [2025-11-24 19:16:42] iteration 1/10 | consumed samples: 64 | learning rate: 1.200000E-05 | global batch size: 64 | lm loss: 1.092887E+01 | loss scale: 1.0 | grad norm: 13.095 | ...
Run 2 - [2025-11-24 19:18:05] iteration 1/10 | consumed samples: 64 | learning rate: 1.200000E-05 | global batch size: 64 | lm loss: 1.092407E+01 | loss scale: 1.0 | grad norm: 13.054 | ...
```

---

**Impact**: High - Blocks reproducibility validation and deterministic testing workflows

**Urgency**: Medium - Deterministic mode works correctly in v0.13.1 but is broken in v0.14.0, forcing users to remain on the older version to maintain reproducibility.