Fix shared memory Table size and accesses reporting in GUPS benchmark

[dmikushin]

Summary

This PR fixes the incorrect Table size and accesses per thread reporting for shared memory tests in the GUPS benchmark, addressing the issues raised by @rkarim2 in #56.

Problem

Previously, when running shared memory tests, the benchmark incorrectly reported:

• Table size showed the global memory working set size (e.g., 4.2 GB), which was irrelevant for shared memory tests
• Number of accesses per thread used the global memory value instead of the shared memory-specific value

Solution

• Table size for shared memory: Now correctly shows the actual total shared memory used across all blocks (grid * n_shmem * sizeof(benchtype))
• Clear details: Added breakdown showing bytes per block × number of blocks for transparency
• Correct accesses count: Uses accesses_per_elem_sh for shared memory tests instead of accesses_per_elem
• Appropriate units: Reports shared memory size in MB (more suitable scale) vs GB for global memory

Example output

Before:

Table size = 536870912 (4.294967 GB.)
Each thread access 8 locations.

After:

Table size = 12288 (0.098304 MB.) [shared memory: 49152 bytes per block x 2 blocks]
Each thread access 65536 locations.

Testing

• Code compiles successfully with make clean && make
• Shared memory tests now display correct memory sizes and access counts
• Global memory tests remain unchanged and work as before

Addresses the confusion mentioned in #56 about the misleading Table size output for shared memory tests.

Note: This is a clean PR with only the relevant changes, created from the latest master branch.

[rkarim2]

Remove the comment on line 497
Otherwise, this looks and works fine.
Thanks @dmikushin