[ESIMD] Fix implementations of block_load(usm, ...) and block_load(acc) (#11797)

1) Fix the big mess in E2E test for block_load(). Test did not really
   check the mask variant. It also used wrong alignments.

2) Fix the comments for USM and ACC block_load implementations. 
3) Minor optimization for ACC block_load functions that do not accept
   the byte_offset operand. We can assume align16 for them.

Signed-off-by: Klochkov, Vyacheslav N <[email protected]>

Author: v-klochkov

sycl/include/sycl/ext/intel/esimd/detail/memory_intrin.hpp

@@ -212,7 +212,7 @@ __ESIMD_INTRIN __ESIMD_DNS::vector_type_t<T, N * __ESIMD_DNS::to_int<VS>()>
 __esimd_lsc_load_merge_bti(
     __ESIMD_DNS::simd_mask_storage_t<N> pred,
     __ESIMD_DNS::vector_type_t<uint32_t, N> offsets, SurfIndAliasT surf_ind,
-    __ESIMD_DNS::vector_type_t<T, N * __ESIMD_DNS::to_int<VS>()> PassThru = 0)
+    __ESIMD_DNS::vector_type_t<T, N * __ESIMD_DNS::to_int<VS>()> PassThru)
 #ifdef __SYCL_DEVICE_ONLY__
     ;
 #else  // __SYCL_DEVICE_ONLY__