[XLA:GPU] Enhance the tiled 0-2-1 transpose algorithm to handle fusion.

Add class TiledParameterInfo to provide information for FusedIrEmitter to read
the content of a tiled parameter from the tile buffer instead of the original
input memory.

Reimplement the tiled 0-2-1 transpose algorithm for copy instructions only in a
more general way so that it can handle both fusion instructions and copy
instructions.

The original tiled 0-2-1 transpose implementation incorrectly used
(tile_size+1) rows for a tile buffer to reduce share memory bank conflicts while
it should be (tile_size+1) column instead. This is a performance issue and is
fixed in the new implementation.

The original tiled 0-2-1 transpose implementation did not generate LLVM alias
meta data for the loads and stores of the tensors. This was due to a bug where
function IrArray::CastToShape miss copying meta data to the new IrArray
object. This is also a performance issue and is fixed in this change.

Modified KernelSupportLibrary to support emitting an if-stmt with a given
branch name prefix.

Add test cases to test the new implementation.

PiperOrigin-RevId: 203310403