Preserve bidirectional unit identity in time-sequential ramp inputs

[zack-dev-cm]

Related to #42 and #39; thanks to tytowatt for identifying the helper and example failures.

Joining prior dispatch to ramp bids on unit alone duplicates generator/load rows, while a stale initial_output produces suffixed columns. Join on (unit, dispatch_type) when available, validate cardinality, and retain authoritative prior output. New units without prior dispatch start at zero; missing ramp values remain available for validation.

The solver regression also establishes the bidirectional-unit contract: both rows need signed net output, generation minus consumption, for the composite ramp constraints. With prior net generation of 4 MW and a 12 MW/hour ramp limit over five minutes, the next bound is 5 MW. Tests cover both signs and crossing zero. Historical unit-level seed output is already net output; single-direction loads retain their positive-consumption convention.

The example passes the three violation costs by keyword, as reported in #42. Scope is limited to the two helpers, their example calls and regressions.

Validation, based on 2d3cef0:

• Original helpers: 23 failing regression cases and four passing controls, including actual two-interval CBC dispatch.
• Hosted Python 3.9/pandas 1.5 and Python 3.12/pandas 3: 83 selected offline tests, two helper doctests and 83 installed-wheel tests pass on each without skips. Installed module bytes match the tested source. These runs combine this PR's source with Prevent bundled CBC crashes with CFFI 2.x on Linux #44's dependency metadata.
• On macOS arm64/Python 3.12.8, all 83 selected tests pass, including the interconnector case.

Linux compatibility: the unchanged upstream interconnector test aborts with CFFI 2.x and passes with 1.17.1. Review the independently scoped #44 first; its installed-wheel checks pass on Python 3.9/3.12. This PR alone retains the existing dependency metadata.

These are contributor-reported results; upstream workflows await approval. Historical cache-dependent benchmarks and complete resolution of #39 are outside the validated scope. The signed-net-output and missing-identity contracts remain explicit review points.