This commit adds a new NIR pass which transforms a cross-invocation
TCS load into a same-invocation load and a subgroup operation.

The subgroup operation is selected based on the TCS input and output
vertex counts and whether the VS and TCS have the same number of
invocations.

Supported subgroup operations are quad_swizze_amd, quad_broadcast
and shuffle.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>