SetCoverLagrangian: stale ThreePhase() declaration, and ComputeLowerBound() segfaults if UseNumThreads() was not called

[jg-codes]

Two small things in ortools/set_cover/set_cover_lagrangian.h, found while writing #5122.
Neither comes from that PR. Checked on main @5a1b660. set_cover_lagrangian.cc is
identical on stable, so both apply to the released wheels too.

1. ThreePhase() is declared but has no definition

set_cover_lagrangian.h:135 declares void ThreePhase(Cost upper_bound);. Nothing defines
it in ortools/set_cover. The three-phase procedure now lives in set_cover_cft.cc as
RunThreePhase, so this looks like a leftover from before CFT moved to its own file.

It is harmless in C++ until someone calls it (link error). Through pybind11 it is worse:
&SetCoverLagrangian::ThreePhase compiles, and the failure only shows up at module import
as undefined symbol: _ZN19operations_research18SetCoverLagrangian10ThreePhaseEd, which takes down the whole set_cover extension.

Suggest deleting the declaration.

2. ComputeLowerBound() needs UseNumThreads() first but has undocumented condition

thread_pool_ is nullptr until UseNumThreads() runs (set_cover_lagrangian.h:60,
:64). ComputeLowerBound() calls the Parallel* methods unconditionally, and they do
thread_pool_->Schedule(...). set_cover_solve.cc:335 calls UseNumThreads() first, so the
solver binary is fine. A fresh SetCoverLagrangian however results in:

lag = set_cover.SetCoverLagrangian(inv)
lag.compute_lower_bound(model.subset_costs, upper_bound)  # SIGSEGV
lag.use_num_threads(1)                                    # fine if called first

The header does not mention the precondition. One could: i) Either default thread_pool_ to one thread in the constructor, or ii) CHECK(thread_pool_ != nullptr) with a message naming UseNumThreads(), or iii) a comment on ComputeLowerBound()`.

Drafted with Claude; reproduced and verified by me.