Add vector-clock lifetimes for concurrent allocation problems

[fpedd]

Lifetimes generalize from scalar interval bounds to vector clocks with one
component per thread, so allocation problems from partially-synchronized
concurrent executions can be expressed directly. Temporal overlap becomes the
happens-before conflict test (neither free happens-before the other's alloc);
scalar problems keep the existing sweep and semantics, and 1-tuples normalize
to scalars. See docs/vector-clock-lifetimes.md for the design.

• Allocation (C++/nanobind) stores TimePoint = int64 | vector, with
  componentwise validation, span accessors, a scalar fast path in
  overlaps_temporally, and ValueError on scalar-accessor misuse
• compute_temporal_overlaps dispatches to the scalar sweep or the pairwise
  vector loop; compute_conflict_degrees counts conflicts per allocation
  (multiplicity-correct under duplicate ids) without marshaling the graph
• get_pressure computes the max-weight antichain of the happens-before order
  via a lower-bounded min flow (weighted Dilworth dual)
• try_linearize synthesizes scalar lifetimes with the identical conflict
  relation when the order is an interval order (no induced 2+2), unlocking
  scalar-only allocators; offsets and conflicts carry over exactly
• allocators declare supports_vector_time; conflict-relation allocators
  (greedy family, naive, order-search) accept vector clocks, minimalloc and
  supermalloc reject them with a precise error, and mixed dimensions are a
  hard error on every path
• run_benchmark skips incompatible allocator/source pairs (decided once per
  variant) and fails with a precise error if a campaign ends up empty
• ConcurrentTilingSource shards guillotine tilings across simulated workers
  with random sync messages; stacked bands keep capacity a provably
  achievable optimum (num_allocations >= num_threads enforced)
• plot_allocation renders one lane per thread, supports memories mixing
  scalar and vector pools, and fixes the pool background color selection
• dump/load round-trip vector times as :-joined components (an omnimalloc
  extension of the minimalloc CSV format)