incorrect handling of undrescores in column names in composite bloom filter

[akuzm]

CREATE TABLE t (
      ts timestamptz NOT NULL,
      a_b int,
      c   int,
      a   int,
      b_c int
  );

  SELECT create_hypertable('t', 'ts');

  ALTER TABLE t SET (
      timescaledb.compress,
      timescaledb.compress_orderby = 'ts',
      timescaledb.compress_index = 'bloom(a_b, c), bloom(a, b_c)'
  );

  INSERT INTO t VALUES ('2024-01-01', 1, 2, 3, 4);

  SELECT compress_chunk(c) FROM show_chunks('t') c;

  -- ERROR:  column "_ts_meta_v2_bloomh_a_b_c" specified more than once

It would be good to fix this for the next release while this feature is still fresh.

[pavanmanishd]

Composite bloom column names join the indexed columns with _, which is not injective when column names contain underscores: (a_b, c) and (a, b_c) both produce _ts_meta_v2_bloomh_a_b_c. The collision extends to the long-name md5-hashed branch since the hash input was the same _-joined string.

Fix: composite names (num_columns > 1) always include a 4-char md5 prefix computed over a \0-separated input, which cannot appear inside an identifier. Single-column names are unchanged. The latest-dev.sql upgrade block walks compression_settings and renames pre-fix composite bloom columns in place (covering both pre-fix short and long branches), preserving pushdown on already-compressed data.

Implementation in #9630.