Bind DB queue names and job payloads as text

[elpete]

DBProvider.push currently asks qb to infer SQL types for every queue name and serialized job payload, even though both fields are text. Bind these two values explicitly as VARCHAR to avoid repeated numeric/date/type inference during native dispatch. Epoch timestamps, delays, attempt bindings, per-call query construction and provider lifecycle hooks retain their existing behavior.

Regression coverage exercises actual native inserts with long Unicode JSON, nested values, literal quote/backslash/newline characters, leading-zero queue names, delayed availability and bigint attempts. This PR is stacked on #32 (fix/db-execution-counter-capacity).

Validation at this candidate:

• A disposable PostgreSQL/BoxLang application passes 39 integration checks covering text persistence, publication, recipient scope and native processing. A separate committed-state probe proves complete rollback on notification conflict, a late deadline and failure of the second native job insertion, followed by successful same-key retry and permanent replay.
• Six alternating pairs of 500 real native pushes, after 100 warmup jobs per provider, give original/candidate medians of 1,704/1,418.5 ms (16.75% lower). The candidate is faster in four of six pairs; the last two are slower. Every sample is rolled back and the committed queue snapshot is unchanged. This measures push cost, not publication p95 or queue throughput.
• The application's installed dependencies are unchanged: only the uniquely owned verifier copy receives this candidate. Candidate runtime hashes and original-provider benchmark source are retained with the results. All disposable resources are independently verified removed.
• At ecec9cf3e36bf01e6e621323f4815789d0d441e8, Lucee 5/6, Adobe 2021/2023/2025, formatting and security checks all pass. The final test uses literal text comparison for queue names.
• The exact narrowed runtime completes all functional checks in a disposable 1,000-offer publication measurement: 2,037 Send statements, 4.864 seconds to the final transaction update and 5.420 seconds through account visibility (including a counter snapshot request). The account gate fails; this single sample is not p95 acceptance. All owned resources are independently verified removed.
• Local CFFormat and git diff --check pass. The application subsequently adopts the exact Git package with all 32 runtime files matched against upstream and root qb unchanged; native installed-package verification remains application-specific follow-up.