Adds prompt lookup decoding (ngram speculation)

[tgaddair]

Adds prompt lookup decoding (ngram speculation) via initialization param --speculative-tokens.

Example:

lorax-launcher ... --speculative-tokens 3

Closes #259.
Related #205.
Related #329.

[lighteternal]

Hi @tgaddair! Many thanks for the new merge 🙏
This new argument --speculative-tokens 3 is not available while running the docker run command in the latest image, is this correct?

Usage: lorax-launcher <--model-id <MODEL_ID>|--adapter-id <ADAPTER_ID>|--source <SOURCE>|--adapter-source <ADAPTER_SOURCE>|--revision <REVISION>|--validation-workers <VALIDATION_WORKERS>|--sharded <SHARDED>|--num-shard <NUM_SHARD>|--quantize <QUANTIZE>|--compile|--dtype <DTYPE>|--trust-remote-code|--max-concurrent-requests <MAX_CONCURRENT_REQUESTS>|--max-best-of <MAX_BEST_OF>|--max-stop-sequences <MAX_STOP_SEQUENCES>|--max-input-length <MAX_INPUT_LENGTH>|--max-total-tokens <MAX_TOTAL_TOK or 'app.py' filENS>|--waiting-served-ratio <WAITING_SERVED_RATIO>|--max-batch-prefill-tokens <MAX_BATCH_PREFILL_TOKENS>|--max-batch-total-tokens <MAX_BATCH_TOTAL_TOKENS>|--max-waiting-tokens <MAX_WAITING_TOKENS>|--max-active-adapters <MAX_ACTIVE_ADAPTERS>|--adapter-cycle-time-s <ADAPTER_CYCLE_TIME_S>|--adapter-memory-fraction <ADAPTER_MEMORY_FRACTION>|--hostname <HOSTNAME>|--port <PORT>|--shard-uds-path <SHARD_UDS_PATH>|--master-addr <MASTER_ADDR>|--master-port <MASTER_PORT>|--huggingface-hub-cache <HUGGINGFACE_HUB_CACHE>|--weights-cache-override <WEIGHTS_CACHE_OVERRIDE>|--disable-custom-kernels|--cuda-memory-fraction <CUDA_MEMORY_FRACTION>|--json-output|--otlp-endpoint <OTLP_ENDPOINT>|--cors-allow-origin <CORS_ALLOW_ORIGIN>|--cors-allow-header <CORS_ALLOW_HEADER>|--cors-expose-header <CORS_EXPOSE_HEADER>|--cors-allow-method <CORS_ALLOW_METHOD>|--cors-allow-credentials <CORS_ALLOW_CREDENTIALS>|--watermark-gamma <WATERMARK_GAMMA>|--watermark-delta <WATERMARK_DELTA>|--ngrok|--ngrok-authtoken <NGROK_AUTHTOKEN>|--ngrok-edge <NGROK_EDGE>|--env|--download-only>