Add Gemma-4 (E2B) text inference sample (torch-neuronx)

[xbill9]

Adds a torch-neuronx/inference sample running google/gemma-4-E2B-it on a single NeuronCore (inf2.xlarge).

Addresses #127.

What it does

A two-graph KV-cache design traced with torch_neuronx.trace:

• prefill — runs the padded prompt, returns logits + per-layer K/V;
• decode — one token vs a fixed MAX-length KV buffer passed as graph I/O, updated with a one-hot masked write (no dynamic index ops).

Token embeddings and per-layer inputs are computed on the host and passed as inputs_embeds / per_layer_inputs, so the traced graph is free of the Per-Layer-Embeddings lookup. Uses eager attention + tanh GELU + final-logit softcap to match the reference.

Validation

Ran end-to-end on a fresh inf2 (Neuron SDK 2.23, torch-neuronx 2.8, transformers 5.13):

• greedy output token-for-token identical to the CPU reference (SEQ_MATCH: True),
• coherent generation, ~45.6 tok/s.

Notes

• Draft pending maintainer feedback (see [Sample request] Gemma-4 (E2B) text inference on Inf2 via torch-neuronx #127): preferred placement (single notebook here vs. a subdirectory with a README documenting the Gemma-4 gotchas), and whether a follow-up TP=2 sample for the 4B/12B variants is in scope for this repo.
• Gated model: the notebook includes the Hugging Face login step and does not bundle weights.
• I'll sign the CLA.