[BUG] signpost does not always report counts correctly

[zoecarver]

Each core only supports 125 signposts so we quickly overflow this in many cases
(especially many ops in a tight loop). When the buffer fills, later signposts are
silently dropped, causing mismatched counts (e.g., 512 cb_reserve but only 384
cb_push for the same with block).

Potential Solutions

1. Increase L1 buffer size - Edit PROFILER_L1_OPTIONAL_MARKER_COUNT in
   tt_metal/hostdevcommon/api/hostdevcommon/profiler_common.h. Cost: more L1 per
   core.
2. Sampling mode - Use modulo in loop to only record every Nth iteration (e.g.,
   if iteration % 64 == 0). Reduces signpost volume while still capturing
   representative timing.