[SYCL][Reduction] Remove atomic64 check for `float` reductions

[pgorlani]

This patch

1. removes the check on aspect::atomic64 for reductions using atomic operations of single precision floating-point values. This solves https://github.com/intel/llvm/issues/6054. However, it's not strictly necessary since also https://github.com/intel/llvm/pull/6429 solves it.

2. adds the possibility of using min and max atomic operations for reductions of double values.

[v-klochkov]

@aelovikov-intel - can you please review this patch as the only active developer of reduction these days?

[aelovikov-intel]

@v-klochkov do we require to run any performance tests before commit?

[aelovikov-intel]

Tag @steffenlarsen , in case you'd like to have a look as well.

[pgorlani]

Thanks for your reviews, @steffenlarsen @aelovikov-intel!

I rebased this PR onto the last refactoring of reductions, it should be fine. The unexpectedly passed test is enabled in https://github.com/intel/llvm-test-suite/pull/1097.

[aelovikov-intel]

Thanks!