perf: Add MergeOrdered IR node with proper sim interleaving hooks

[Benjscho]

Addresses #2768

Replace the sliced!-based merge_ordered implementation with a dedicated HydroNode::MergeOrdered IR node that has distinct codegen for production and simulation.

Production codegen:

• Emits union() DFIR operator (both inputs pulled in same stratum)

Simulation codegen (two modes):

• Bounded/tick: MergeOrderedHook (inline) buffers both batches and generates a valid interleaving via boolean coin flips, preserving per-input order. Explores C(a+b, a) states instead of (a+b)!.
• Unbounded/top-level: TopLevelMergeOrderedHook releases one element at a time from the front of either input queue, allowing feedback cycles to deliver elements between releases.

API changes:

• Removed NoTick bound from merge_ordered — works on any Location and Boundedness now. When bounded, the nondet is still explored in sim.
• State space reduced from 26 to 6 instances in the existing test (2 elements per input: C(4,2) = 6 valid interleavings).

Tests added:

• sim_merge_ordered: ordering preservation assertion (invalid interleavings like [2,1,3,4] are never produced)
• sim_merge_ordered_one_empty: pass-through when one input is empty
• sim_merge_ordered_cycle_back: feedback cycle where a cycled-back element arrives before elements on the other input
• sim_merge_ordered_delayed: delayed element on one input
• deploy_merge_ordered_delayed: production localhost test with timing