✨ Use native QDMI multi-program jobs in PennyLane workflows

[burgholzer]

🤖 AI text below 🤖

Elevator pitch

PennyLane produces batches through broadcasting, measurement grouping, differentiation, and shot vectors. Native batching should improve these workflows without undoing concurrent single-program execution.

Current state

python/mqt/core/plugins/pennylane/device.py maps preprocessed tapes to jobs. #2349 submits individual jobs before waiting and preserves ordered results and best-effort cleanup.

Work and decisions

• Build on the latest concurrent execution path, not the older sequential code in ✨ Adopt exact QDMI 1.4 payload contracts across Core #2226.
• Use native multi-program jobs only for compatible format, shot, and job-option groups.
• Restore tape, measurement-group, broadcast, and shot-vector order after indexed decoding.
• Keep concurrent single-program submission as the unsupported-batch fallback.
• Preserve counters, batch timing, error propagation, interruption handling, and best-effort cancellation without masking the original error.

Dependencies and release boundary

Parent: ✨ Adopt native multi-program QDMI jobs across Core and SDK plugins. Requires ✨ Extract QDMI multi-program Client and bundled-device support. Preserve #2349 and any successor fixes.

Acceptance and validation

• Tests compare native and fallback results for ordinary tapes, broadcasts, non-commuting groups, parameter-shift batches, and shot vectors.
• Failure and KeyboardInterrupt cleanup preserves the first error and attempts cancellation of submitted work.
• Unsupported batching does not regress asynchronous overlap.
• Run focused PennyLane/DDSIM tests, supported/minimum sessions, and full lint.

Non-goals

Requiring the MQT compiler for batching; payload capability design; changing supported gate decomposition or sampling semantics.

Provider compatibility evidence

Contribute an end-to-end example to #2359 using the matching provider implementation. Verify ordered SDK results and lifecycle behavior for both supported native batches and concurrent single-program fallback. Do not infer aggregate support from a provider SDK's batch convenience API.