[AMDGPU] Add islands to StructurizeCFG
Handle an "island", i.e. a block whose terminator is not rewritable
(e.g., a callbr) in this context and considered immutable from this
pass's perspective.
The island is left untouched; its edges are split into forwarders that
converge at a new Flow block (ExitFlow). The callbr's runtime target
choice is recovered as a per-target i1 "sel" phi and re-dispatched by a
structured ladder of 2-way branches:
callbr
/ | \
fwd_0 fwd_1 fwd_2 forwarders (intermediate target blocks)
\ | /
ExitFlow (Flow_0) -- sel_0? --> real_0
| else
Flow_1 --------- sel_1? --> real_1
| else
real_2
or Flow_2 (if real_1 == real_2) ------> real_2 (shared last target)
