Implement the public API audit

[fpedd]

Land the rulings of research/api-audit.md across the Python and C++
surface: one vocabulary, flat functions at the boundary, noun-phrase
queries, flat allocator kwargs, renamed entry points, and omni as the
default allocator.

Vocabulary (R1): "conflict" names the happens-before constraint,
"overlap" stays reserved for realized rectangle collisions.
Allocation.overlaps_temporally becomes conflicts_with; the C++ tree
renames in lockstep (ConflictMap, conflict_degrees,
SortKey.CONFLICTS/DURATION/START/END); the conflict map is now total,
so the Python wrapper's .get() patching dissolves.

Binding surface (R4-R6, R14, R16, R17): the boundary carries data types
and flat functions only. The stateless *AllocatorCpp classes become
*_place functions with default-free parameters - Python owns every
policy default. Failures use builtin exceptions only: budget exhaustion
raises RuntimeError, a missing optional dependency raises ImportError -
BudgetExceededError, OptionalDependencyError, the Option unions, and
the Python-side patching all dissolve. try_solve_many binds the search
switches flat with max_nodes: int | None (None = unbounded); Solution
reshapes to .allocations/.peak; clique_cap and Allocation.with_kind
delete codebase-wide. FirstFitPlacer rejects repeated order indices
alongside out-of-range ones.

Query names (R2, R9, R13, R15): pure queries are noun phrases -
conflicts, conflict_degrees, pressure, closure_pressure,
placement_pressure, with *_per_allocation variants on both sides of the
boundary. The top level narrows to conflicts, pressure, and
try_linearize; the full family lives in omnimalloc.analysis.
greedy_base's duplicate peak/conflict sweeps dissolve into the analysis
queries; conflict_degrees counts scalar timelines in O(N log N) via two
binary searches, skipping the pairwise sweep; README and examples adopt
import omnimalloc as om.

Allocator ergonomics (R7, R8, R11, R14): flat keyword-only constructor
kwargs replace the four config dataclasses (scikit-learn style), with a
shared self-describing repr; common/validation.py and
common/parallel.py spell the numeric precondition checks and the
num_threads: int | None = None convention once; registry names strip
the class-role token as a suffix only ("greedy_by_size");
GeneticAllocator's num_generations becomes max_generations. The 8 *Cpp
greedy twin classes are removed - two spellings of one code path
measuring nothing. Flat plain-typed kwargs make every allocator
picklable, so allocate_parallel submits configured variants directly
instead of rejecting them.

Entry points (sections 5 and 6): run_allocation becomes
omnimalloc.allocate(pool); validate_allocation(entity) -> None is
raise-only; plot_allocation(entity, path=None) displays or saves, and
plot_benchmark(artifact, path=None) matches; dump_allocation becomes
save_allocation (module io.py) and round-trips placements through the
new offset column, written whenever any allocation is placed;
the minimalloc benchmark source loads through io.load_allocation;
Memory.size becomes capacity; BufferKind becomes AllocationKind, C++
enum included.

DEFAULT_ALLOCATOR becomes "omni": never worse than greedy_by_size on
any benchmarked case (geomean peak ratio 1.0972 vs 1.1675), certified
optimal on three of seven, vector-clock capable, and the same order of
runtime.