Prebuild and slim the dependency planner

[smcio]

the dependency planner currently clones tx’s unnecessarily and builds the conflict graph synchronously inside the transaction handling pipeline. this was improved by:

• replace full transaction cloning with compact, deduplicated planner account snapshots
• prebuild dependency plans alongside account loading and join only when TxLoop needs the result
• use a transitive-reduced graph and compact CSR representation to reduce work and allocations
• handle unresolved address-table metadata conservatively while retaining the Lightbringer fallback
• fix fallback batch tracking so later conflicting accesses cannot be scheduled into an earlier batch
• add planner preparation/build/wait/dispatch metrics, viewer support, regression tests, fuzz coverage, and benchmarks

Results:

• SlotReplay: 18.186 ms → 17.309 ms (4.8% lower)
• non-empty replay cost: 17.633 µs/tx → 16.434 µs/tx (6.8% lower)
• DependencyPlannerBuild: 0.851 ms → 0.220 ms (74.1% lower)
• TxLoop: 2.119 ms → 1.235 ms (41.7% lower)