Organize the dot op emitter a bit better; NFC

This CL pays down some technical debt around dot_op_emitter's organization.  It

 - Extracts out a GetDotImplementationStrategy that decides the lowering
   strategy for a dot operation.  Earlier this logic was scattered through the
   lowering code.

 - Removes calls to PotentiallyImplementedAsEigenDot.  This is an implementation
   detail for dot_op_emitter and exposing it like this has aged poorly in the
   face of the various tiled LLVM IR lowerings we now have.  Instead, we expose
   two more specific hooks: DotOperandsAndResultMustHaveRowMajorLayout and
   DotImplementationCanHandleTranspose.

 - Renames xla_enable_experimental_llvm_ir_gemm to
   xla_force_enable_experimental_llvm_ir_gemm since its behavior is that it
   enables the tiled LLVM IR GEMM even when it does not look profitable.

Note: In parallel_task_assignment the call to PotentiallyImplementedAsEigenDot
was redundant since we already check for non-loop fusions and dots.
PiperOrigin-RevId: 228276282