Use CHS published chart-datum offsets for Canadian stations instead of computed LLWLT

While validating #110, I benchmarked our datums against the CHS API for the 186 quality-accepted -can-meds stations. CHS metadata publishes each station's chart-datum offset, so we can compare their real chart datum against what we compute.

The result: computed LLWLT is a worse stand-in for CHS chart datum than LAT was.

The cause is not a bug in the LLWLT math. CHS chart datum is a frozen benchmark: it was established decades ago and is deliberately not recomputed, while our LLWLT is the textbook construct derived from today's constituents over the current epoch. Decades of relative sea-level change leave the operational datum much closer to modern LAT than a fresh LLWLT calculation, and occasionally below it. No constituent-derived datum can reproduce a fixed benchmark.

The saving grace of LLWLT is direction: our zero now sits about 0.17 m above CHS's, so predictions understate water depth, which is the safe error for navigation. The old LAT fallback overstated depth by about 0.09 m.

Proposal

For -can-meds stations, fetch the station's chart-datum offset from the CHS API (api-sine.dfo-mpo.gc.ca, the same metadata endpoint the neaps benchmark uses) during import and reference chart_datum to that published value instead of the computed LLWLT. That makes the error zero by construction, since we would be using the agency's own number.

Open questions:

• Where to store it. Probably overwrite the LLWLT value with the official offset (keeping the name), or add a separate datum entry so the computed value stays available.
• Terms of use for the CHS API data in a redistributed dataset need a check before we bake values into the repo.
• The same reasoning applies to any agency with a fixed benchmark. Japan (NLLW) and China (TLT) may have the same drift between the official surface and our computed one; worth checking if an equivalent API or table exists before trusting the computed values there.

Benchmark data and method are in the PR discussion for #110; the comparison used cached CHS metadata from neaps/benchmarks.

Start with the PR discussion for #110 and the cached CHS metadata in neaps/benchmarks, then trace the -can-meds import path and the neaps benchmark's CHS metadata endpoint at api-sine.dfo-mpo.gc.ca. Determine how the published chart_datum should be stored, verify redistribution terms, and confirm that imported Canadian stations use the published offset rather than computed LLWLT.