Convert to Triton Punica kernels

[tgaddair]

• Add new Triton-based Punica kernels from vLLM that are faster and more flexible than existing Punica kernels
• Allow max_batch_prefill_tokens < max_input_length when chunked_prefill is enabled (chunk input sequence)
• Enable profiling when LORAX_PROFILER_DIR is set
• Fixed incompatibility between prefix caching and speculative decoding (Fix prefix caching + speculative decoding huggingface/text-generation-inference#2711)
• Fixed multi-Medusa to work with CUDA graphs
• Added Triton kernels to optimize batch preprocessing (feat: add triton kernels to decrease latency of large batches huggingface/text-generation-inference#2687)
• Trace preloaded adapter layers during initialization, instead of generic model defaults
• Remove flume (feat: remove flume huggingface/text-generation-inference#1184)
• Update Axum and router dependencies (Upgrade to Axum 0.7 and Hyper 1.0 (Breaking change: disabled ngrok tunneling). huggingface/text-generation-inference#1959)
• Fixed race condition causing stopped entries not to be removed from the cached batch

Breaking changes:

• Dynamic adapter loading and preloaded adapters can no longer be used together.

[tgaddair]

@noyoshi @magdyksaleh not sure if this is needed and we need to add this back? Let me know.

[magdyksaleh]

yes can we please keep it

[tgaddair]

Looks like this was only enabled when using ngrok tunneling, which is unsupported at the moment. Are we using ngrok today?