Unified pipeline for models & support phi3 model

[guoqingbao]

This PR provides a unified pipeline for models and adds the phi3 model under the unified pipeline. The major changes are:

1. The LLaMaPipeline is removed and replaced with DefaultPipeline (and default loader), which can serve different models such as LLaMa and phi3.
2. The phi3 model is added and tested, achieving around 100 tokens/second generation speed on A100 (BF16).
3. Configuration for models is simplified.
4. The padding strategy is optimized.
5. ReadMe is revised to reflect the recent changes.

More models are expected to be added using the DefaultPipeline.

Mixed precision is used for the phi3 model because I found rope and rmsnorm require at least FP32 for long sequence generation (e.g., over 2000 tokens prompt).

Command line to run Phi3 3.8B chat service

cargo run --release -- --port 2000 --weight-path /home/phi3-3.8b/ phi3 --repeat-last-n 64

Please ignore the previous commit messages :)

Mention #44

[EricLBuehler]

Thank you!