fix: include torch compile flag

fix: support torch compile via fake tensors

fix: add quickstart and avoid autotune when no cuda

feat: support shared experts layer and tests

fix: extract expert device mesh for group from unused prehook

fix: prefer using passed parallel group

fix: adjust layer params in source

fix: improve expert parallel implementation and refactors

fix: add parallel forward functional logic

fix: fully vendor stk and fix imports

fix: vendor stk decorators

feat: add functional version of moe class and layer

fix: prefer relative imports

feat: vendor grouped gemm

feat: validate build with original test suite

feat: initial port of megablocks to builder format