Fix YAML-loaded ParameterSet lengths breaking mixed-unit arithmetic (#219)

[ad-cqc]

Summary

• Closes YAML-loaded ParameterSet lengths break mixed-unit arithmetic #219.
• Unitful.uparse (used inside ParameterSetYAMLExt) returns FreeUnits quantities whose promotion target is m. The rest of DeviceLayout uses ContextUnits from DeviceLayout.PreferredUnits (target nm, or μm under PreferMicrons). Mixing the two through + — e.g. inside layer_z in technologies.jl — triggers the upstream Unitful mixed-context promotion bug (JuliaPhysics/Unitful.jl#845) and fails with MethodError: no method matching (::FreeUnits{(m,), 𝐋, nothing})().
• Add a DeviceLayout.uparse(str) that resolves length symbols against PreferredUnits first (so parsed lengths carry the package's ContextUnits from the start) and falls back to Unitful for non-length symbols (s, Hz, fH, …).
• Use DeviceLayout.uparse inside ParameterSetYAMLExt._parse_units! instead of post-parse re-contexting.

After this PR YAML-loaded lengths enter the system already wrapped in ContextUnits matching PreferredUnits.UPREFERRED, so subsequent mixed-context arithmetic (level_z, layer_z, …) avoids the buggy promotion path entirely.

Why a new uparse instead of post-parse re-contexting

The earlier iteration of this PR re-wrapped each parsed Quantity in a ContextUnits aimed at UPREFERRED. That worked but only for length leaves the helper recognized; compound expressions (1/μm, products, etc.) needed extra special-casing. Routing through Unitful.uparse(str; unit_context=PreferredUnits) lets the parser itself produce ContextUnits for any expression involving length symbols, and falls back to Unitful for everything else — no post-processing needed.

Test plan

• Pkg.test() — test/test_parameter_set.jl passes 188/188 (+2 new assertions).
• Existing YAML round-trip tests still pass (serialized form unchanged: string(ContextUnits unit) == "μm").
• New regression cases:
  • DeviceLayout.uparse("150μm") |> unit isa ContextUnits
  • Compound 1/μm round-trips through the YAML loader as a Quantity.
  • Reproducer of the original level_z failure path: sum(t_chips) + t_gap - 1nm followed by another + 1nm no longer throws.

Notes

• Non-length quantities (frequencies, times, kinetic-inductance fH, …) pass through unchanged via the try/catch fallback.
• New public surface: DeviceLayout.uparse (not exported). Useful anywhere external string input needs to land in the package's promotion context — YAML, JSON, REPL helpers.
• PreferNoUnits (UPREFERRED == NoUnits): PreferNoUnits re-exports Unitful.μm (bare FreeUnits), so DeviceLayout.uparse returns identical values to Unitful.uparse — no behavior change in that mode.

[gpeairs]

Quick pre-review note -- BSpline test failures are unrelated (caused by a dependency update). I'm relaxing the tolerances in #217.

The bug is a bit surprising, since the whole point of ContextUnits is that you should be able to use them with FreeUnits and the result gets promoted based on the context. There is one bug I know of JuliaPhysics/Unitful.jl#845 but that's for two different ContextUnits (specifically same unit, different context). I'm not reproducing the error with the given snippet, so maybe you actually hit that bug downstream somewhere. Either way this looks like the correct workaround.

[ad-cqc]

You are right. The code where I spotted the error leads to _promote_unit call in JuliaPhysics/Unitful.jl#845 where units from different contexts getting promoted incorrectly.

So the fix it the same use preferred units from the default context and promote YAML unitful values to the type with new context.

[ad-cqc]

I reworked PR by introducing PreferredUnits aware uparse, so any length parameters would be associated with correct internal context.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!