Improvements Dispatch Generator - MLIR Source Generation and Defining Dispatches (#12792)

This PR is in series of smaller PRs to iteratively improve IREE dispatch
profiler. The improvements captured in this PR are as follows:
- `EmitSourceMLIR` instead of `EmitMatmulSourceMlir`: Once the
`operation` and `configuration` are properly abstracted out, there was
very little matmul-specific code for emitting dispatch (operation +
configuration). Thus, moved EmitSourceMLIR into its own structure which
can be potentially used by new dispatches.

- Dispatches are **generated** and **profiled** from a predefined
shapes, datatypes, and tuning configurations. Both generation and
profiling loads the predefined dispatches for a specific data type in a
`Manifest` object. The definitions of dispatches are operation- and
device-specific. Thus, abstracted out into its own structure `class
MatmulGenerator`. Which can be loaded using `Manifest.load()`.

- Additionally, `class MatmulGenerator` executes a few simple checks to
skip dispatches that a user might naively define but not supported
because of architectural constraints (e.g. shared memory capacity) or
unsupported codegen path (e.g. unaligned matmuls on CUDA tensor core
backend).