Fix Issue #140: GGML Assert Batch Size Failure with Large Prompts

[Michael-A-Kuykendall]

This PR fixes the GGML assert failure when processing large prompts that exceed the hardcoded batch size limit.

Problem

Issue #140 reported GGML_ASSERT(n_tokens_all <= cparams.n_batch) failures when processing large prompts, particularly with DeepSeek-R1 models that have extensive system messages.

Root Cause

The batch size was hardcoded to 2048 tokens, causing assertion failures when context length exceeded this limit.

Solution

Implemented adaptive batch size calculation that scales based on context length:

• Minimum batch size: 2048 tokens
• Maximum batch size: 8192 tokens
• Scales dynamically to ensure batch size >= context length

Changes

• Added calculate_adaptive_batch_size() method in src/engine/llama.rs
• Modified context creation to use adaptive batch size instead of hardcoded 2048
• Added comprehensive regression tests in tests/regression/issue_140_ggml_assert_batch_size.rs

Testing

• All existing tests pass (566 total)
• New regression tests validate batch size calculation for various context lengths
• Release gates pass: build, CUDA, packaging, binary size, test suite, documentation

Impact

This fix prevents server crashes when processing large prompts while maintaining memory efficiency through adaptive sizing.