[shortfin llm] Separate server configs and exported model configs

[renxida]

Currently, the prefix sharing algorithm is configured from config.json, which makes it necessary to edit it every time we export a model before server would work.

config.json should be only used for exported config keys from sharktank.export_paged_llm_v1

server.py should take additional options either from a separate config file or from the commandline arguments.

Ideally we figure out some project-wide standards for configuration management (e.g. json / yml files in ~/.shortfin with cmdline options to override them).