[Disco] Switch to build-time sharding and enable FT quantization

[masahi]

Building on @Lunderberg's work in mlc-ai#1096, we are now switching to build-time sharding in mlc_serve. Runtime sharding is no longer supported, and when building a model you must add --use-presharded-weights. You also need the latest contrib-vllm.

Build-time sharding also lets us support FT quantization with Disco. Now weight preprocessing is applied after sharding. It has been confirmed to work on 7B and 13B, with both q4f16_ft and q8f16_ft, with --num-shards=2. Other configs will be tested later.

[Lunderberg]

As an addendum, the build.py script must be run in two steps, once with --convert-weight-only and once with --build-model-only. This is implemented in this check, and is due to the same parameter size handling described here.