enh(pmc_l2_analyzer): support rocpd SQLite DB input (rocprofv3 >= ROCm 7.x)

[Arist12]

Problem

rocprofv3 --pmc on ROCm >= 7.x writes PMC counter results to a SQLite
database (pmc_*_results.db) rather than a flat CSV file. The previous
pmc_l2_analyzer.py only accepted CSV input, so any analysis on a ROCm 7.x
system required a manual intermediate step: open the DB, query the four rocpd
tables, export to a compatibility CSV, then run the analyzer.

Solution

Add native rocpd SQLite DB support alongside the existing CSV path.

New internal function _load_counters_from_db(path, kernel):

• Opens the DB read-only via sqlite3.connect("file:...?mode=ro", uri=True).
• Discovers the UUID-suffixed table set by scanning sqlite_master for
  rocpd_pmc_event_<uuid>. The UUID suffix is opaque and changes between
  runs, so table names cannot be hard-coded.
• Joins the four required tables:

  rocpd_info_kernel_symbol_<uuid>  (id, name)
  rocpd_kernel_dispatch_<uuid>     (id, kernel_symbol_id)
  rocpd_info_pmc_<uuid>            (id, name)
  rocpd_pmc_event_<uuid>           (id, pmc_id, dispatch_id, value)
  

• Filters dispatches by --kernel substring on the kernel symbol name.
• Returns the same (agg, dispatches) pair as the CSV path so the rest of
  main() is unchanged.
• Prints a warning if any required table is missing rather than crashing.

load_counters() dispatches on file extension:

• .db → rocpd path
• anything else → CSV path (unchanged behavior)

Multiple files of mixed types can be passed in one invocation (the typical
workflow: two separate --pmc runs produce two DBs):

python pmc_l2_analyzer.py pmc_l2_results.db pmc_ea_results.db \
    --kernel pa_mqa_logits_fp4_kernel_0

Also removes the pre-existing dead assignment req = agg.get("TCC_REQ_sum", 0)
that was never used.

Usage comparison

# ROCm < 7.x  (CSV output, unchanged)
python pmc_l2_analyzer.py pmc_l2_counter_collection.csv pmc_ea_counter_collection.csv \
    --kernel pa_mqa_logits_fp4_kernel_0

# ROCm >= 7.x  (rocpd DB output, new)
python pmc_l2_analyzer.py pmc_l2_results.db pmc_ea_results.db \
    --kernel pa_mqa_logits_fp4_kernel_0

# Mixed
python pmc_l2_analyzer.py pmc_l2_results.db pmc_ea_counter_collection.csv \
    --kernel pa_mqa_logits_fp4_kernel_0

Testing

Eight unit tests using tempfile mock DBs and CSVs:

1. CSV path with kernel filter — unchanged behavior.
2. CSV path with no filter — unchanged behavior.
3. rocpd DB with kernel filter — correct counter aggregation.
4. rocpd DB with no filter — returns all dispatches.
5. Mixed DB + CSV in one call — counters are summed correctly.
6. DB with non-trivial UUID suffix — UUID discovery works.
7. Empty DB (no rocpd tables) — returns empty gracefully, no crash.
8. Kernel name not found in DB — returns empty gracefully.

All eight pass.