[SYCL-MLIR] Add KernelDisjointSpecialization pass in pipeline

[whitneywhtsang]

Verified that with this PR and -D__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__, SYCL-MLIR is able to perform scalar replacement on reduction loop on the SYCL-Bench workloads identified before, even with the fix in #9055.

O3	Previously measured gains	Only PR9187	PR9187+PR9055
2mm	12%	12%	12%
3mm	13%	12%	11%
covariance	50%	50%	0%
gemm	12%	12%	11%
gramschmidt	14% + 18% (LICM versioning)	33%	33%
syrk	5%	5%	5%

=> No performance regressions with just this PR.
=> Lost covariance 50% gain after specializing the function, because 2 of the 3 accessors actually overlap!

auto data = data_buffer.get_access<access::mode::read>(cgh);
auto symmat = symmat_buffer.get_access<access::mode::discard_write>(cgh);
auto symmat2 = symmat_buffer.get_access<access::mode::discard_write>(cgh);

Not sure why it is written that way, with the simple source code change below (which remove symmat2 and always use symmat), the 50% gain is recovered.

+++ b/polybench/covariance.cpp
@@ -110,7 +110,6 @@ public:
-                       auto symmat2 = symmat_buffer.get_access<access::mode::discard_write>(cgh);
@@ -122,7 +121,7 @@ public:
-                                       symmat2[{j2, j1}] = symmat[{j1, j2}];
+                                       symmat[{j2, j1}] = symmat[{j1, j2}];

If we want to get the gain without source code change, then we need to version by only checking if symmat and data overlap, which we need the context, we may want to perform loop versioning in detect reduction pass.
Notice the inner loop with the opportunity doesn't use symmat2:

for(size_t i = 1; i <= N_; i++)
  symmat[{j1, j2}] += data[{i, j1}] * data[{i, j2}];

Note: A number of KernelFusion test cases are moved to xfail, as accessors cannot be internalize when they are used by a ptrtoint operation. (#9188)

[etiotto]

"If we want to get the gain without source code change, then we need to version by only checking if symmat and data overlap, which we need the context, we may want to perform loop versioning in detect reduction pass."

Yes I think we need to version the reduction loop to get the opportunity.

[whitneywhtsang]

After allowing functions indirectly called by GPU kernel in #9194, there are more KernelFusion tests failed due to unable to internalized. There are a total of 15 KernelFusion test cases moved to xfail.