[XLA:GPU] Refactor the 0-2-1 tiling implementation to support more use cases.

The current implementation of 0-2-1 tiling only handles two cases, a single
kCopy instruction and a kLoop fusion instruction. This change adds class
TilingScheme to support more general tiling methods that consist of tiles and
block of tiles. It also restructures the existing tiling implementation to
generate the kernel based on a TilingScheme object, a few function objects that
provide hooks for HLO specific code generation and other parameters. This
change supports our next step in extending tiling for reduction.

PiperOrigin-RevId: 221393383