[XLA] Avoid implicit layout change for non-layout-changing instructions.

Add routine InstructionRequiresInputLayoutEqualToOutputLayout to layout
assignment and use it to decide whether the output layout of an instruction
need to be propagated to the operands of the instruction during layout
assignment. This also fixes the performance issue exposed in b/112646847 where
a kSlice instruction with implicit layout change is fused to a multiple output
fusion for reduction and causes memory throttle.

Add test cases.

PiperOrigin-RevId: 210622658