Feat: hybrid CPU/GPU n-gram hashing with automatic path selection

[Playmaker3334]

Summary

This PR adds GPU support to the Engram demo and fixes bugs that prevented it from running on CUDA devices.

Bug Fixes

1. Device mismatch - torch.from_numpy() in Engram.forward() always created CPU tensors, causing runtime errors when model was on GPU
2. Index out of bounds - CompressedTokenizer._compress() had no bounds checking, could crash with certain input_ids

New Features

HybridNgramHashMapping

Replaces NgramHashMapping with a proper nn.Module that supports both CPU and GPU:

if input_size < self.gpu_threshold:
    # NumPy path - lower overhead for small inputs
else:
    # PyTorch path - better throughput for large inputs

Key changes:

• Multipliers stored as register_buffer() for automatic device transfer
• _hash_gpu() uses torch.bitwise_xor instead of numpy
• Configurable threshold via EngramConfig.gpu_threshold

CompressedTokenizer

• compress_cpu(): fast numpy path with bounds checking
• compress_gpu(): lazy tensor initialization, tracks device

Benchmark Suite

New files for validation:

• benchmark.py: measures latency across configs
• test_correctness.py: verifies numerical equivalence

Benchmark Results

Metric	Value
Mean speedup	1.02x
Median speedup	1.01x
Max speedup	1.09x
Memory delta	-0.03%

Tested on NVIDIA GPU with batch_size=[2,4,8], seq_len=[128,256,512]

Backward Compatibility

• Original engram_demo_v1.py unchanged
• All original APIs preserved in optimized version
• Numerical outputs match within rtol=1e-4, atol=1e-5