fix: resolve composite event columns via faers_get, support all MedDRA hierarchy levels

[MadDERt]

Description

faers_phv_composite() / faers_phv_signal_composite() were broken for
MedDRA hierarchy event types and bypassed the duckdb backend. This PR fixes
the event-column resolution and hardens the input validation.

The core fix: .identify_event_set_patients() now resolves event columns
via faers_get(object, "reac"), which attaches the MedDRA hierarchy columns
on both the memory and the duckdb backends — the same resolution path used
by faers_counts() and faers_phv_scan().

Problems fixed

1. Hierarchy event types crashed — .event_type = "soc_name" aborted
   with object 'soc_name' not found on both backends, because the raw
   reac table lacks hierarchy columns. Now works for any standardized
   reac column (pt, meddra_code, meddra_pt, and all MedDRA hierarchy
   columns such as soc_name / hlgt_name), with a helpful error message
   listing the available columns.
2. The duckdb backend was bypassed — db_collect(con, "reac") pulled the
   whole table into memory; removed in favor of faers_get().
3. Dead code that crashed on duckdb-backed .full objects — an unused
   faers_counts() call went through counts_db() and failed with a SQL
   binder error; removed.
4. De-duplication is now enforced (faers_dedup() required), consistent
   with faers_phv_scan().
5. Undeclared assertthat usage replaced with the internal assert_()
   helper (same style as scan).
6. A no-op setnames() line was removed.

Testing

• New tests/testthat/test_signal-eventset.R (the file did not exist):
  26 assertions — soc_name/hlgt_name regression, character & function
  event sets, 2×2 conservation, .object2 comparison mode, the
  de-duplication guard, and memory ↔ duckdb identical() parity
• Full suite: 361 passing; the only failures are the pre-existing
  test_meta.R network/cache environment items (unchanged baseline)
• R CMD check: no new issues introduced

Notes

• Behavior change: de-duplicated objects are now required (consistent with
  faers_phv_scan()); raw pt mode output is otherwise unchanged
• Follow-up (not in this PR): duckdb SQL pushdown so .full composite calls
  filter in-database instead of materializing the table

[ShixiangWang]

@MadDERt 如果是ai工作生成的，确认代码你都自己阅读过，然后check过了你自己可以直接合并。

后面的pr都这样的标准处理