xe: enable resuable_dispatcher_t to handle all buffer layouts

[rjoursler]

JIC anyone wants to review this over the holidays. There is still a little bit of work remaining to ensure there are no functional or performance regression, but the core changes are now implemented.

This PR modifies the generation and encoding used by reusable_dispatcher_t to significantly increase flexibility. This increased flexibility is then used to make the reference eltwise kernel 100% completely reusable. The key changes in this PR:

• Switch to directly encoding expressions for calculations - this enables better compression of the expressions, so that more buffers can be registered, to encode expressions unrelated to buffer offsets (i.e. gws_overflow and gws_in_padding), and to optimize expensive computations (i.e. the addition idiv).
• Switch buffer offsets computation to be in terms of the outer dimensions i.e. sum(outer_dim_idx * outer_dim_stride) + offset(get_inner_dim()). This enables properly inclining constants associated with blocked layouts, which avoids expensive divisions.

Beyond that, there are a few other (but more minor) optimizations at play

• Switch named buffer encoding to a bitset. This reduces the overall structure size as we do not need to store buffer names and normalizes the structure layout which could prevent non-determinism due to ordering.
• Use 32-bit values in the runtime params when possible. This reduces the necessary data transfer to the kernel, and also avoids emulated 64-bit arithmetic.

[rjoursler]

make test
disable test_device_cpu