[OpenCB Project Fix] Fixes #26005: ClassTag synthesis for tvar with chained tvar bound

[soronpo]

When synthesizedClassTag is asked for a ClassTag[E1] where the only constraint on E1 is E1 <: E and E is itself an unsolved type variable, the previous logic could not reach a ground instantiation:

• isGroundConstr(fullUpperBound(E1.origin)) was false (the bound contains an unsolved tvar, namely E).
• isFullyDefined(tp, ForceDegree.all) would default E1 (and E) to their lower bound (Nothing), producing a bottom type the synthesizer rejects, yielding "No ClassTag available for E1".

The fix layers two new behaviours on top of the existing path:

1. Try isFullyDefined(tp, ForceDegree.failBottom) first. This succeeds whenever the surrounding inference has pinned a real answer (preserving the i23611 backward-compatible behaviour) but bails when the only choice would be a bottom default.
2. If failBottom bails and the chain of unsolved tvars in the upper-bound direction reaches a ground type (e.g. E <: Throwable), instantiate the chain's parent tvar from that ground bound and then instantiate the original tvar from the now-ground upper bound.
3. Fall back to ForceDegree.all last, preserving original behaviour when neither chain-walking nor failBottom apply.

The chain walker (forceGround) handles the AndType case where the constraint is recorded as TypeVar(E) & TypeParamRef(E) (logically a single tvar) and recurses through TypeVar/TypeParamRef forms, with a seen set to avoid infinite loops on bidirectionally-equated tvars.

Fixes #26005

How much have you relied on LLM-based tools in this contribution?

Extensively, but first want to see if CI is green. This is a long-standing issue blocking a project on the community build.

How was the solution tested?

New automated tests (including the issue's reproducer, if applicable)