feat: add rolling cache turn pruning for ChatSampler (Issue #675)

[Paramveersingh-S]

Description

This PR addresses the Context Exhaustion issue outlined in Issue #675, specifically focusing on the ChatSampler crashing when long multi-turn conversations exceed the static 4096-token cache_length.

Since JAX arrays are statically compiled and dynamic jnp.roll sliding-window operations introduce significant compilation and latency overheads, this PR solves the issue at the orchestration layer by implementing Context Window Management (Turn Pruning) directly inside gemma/gm/text/_chat_sampler.py.

Key Changes

• Automated Context Pruning: Added a _prune_context_to_fit mechanism to the ChatSampler.chat method. Before triggering the SamplerLoop, it calculates if used_cache + new_prompt_tokens + max_out_length > cache_length.
• Eviction Strategy: If the context overflows, the sampler strategically pops the oldest User/Model conversation turn pair from self.turns while explicitly preserving the initial System prompt (if present).
• Media History Tracking: Introduced history_images and history_audio properties to the ChatSampler. This ensures that when the context is pruned, the sampler can safely flush the static last_state KV Cache and execute a full re-prefill using the dynamically retained multimodal history without dropping user-provided media from active turns.
• Unit Testing: Added _chat_sampler_test.py to statically verify the eviction constraints.

Fixes #675.