[AutoDeploy] Detach sharding pattern matcher from pattern executor

[greg-kwasniewski1]

Sharding detection (TP, EP, BMM) should create a "sharding config", instead of directly modifying the compute graph, they should parse the graph read-only.

The config should be then processed by a separate pattern executor that apply graph transformations provided.

This way, the autosharding logic could be bypassed/overwritten by manually specifying the sharding config.