Replace msm execution contexts with cudaq::dem_from_kernel in app examples

[bmhowe23]

The surface_code-2 and surface_code-3 C++ app examples built their detector error models by running the demo kernels under the "msm_size"/"msm" execution contexts and hand-assembling the detector and observable matrices
from the raw measurement syndrome matrix. The cudaq::dem_from_kernel API supersedes that: the demo kernels now annotate themselves with cudaq::detector/detectors and cudaq::logical_observable when a declare_detectors flag is set (detectors pair each round against the previous one, with the lock-in round as the reference; prep0's deterministic first-round Z stabilizers are declared as singles where the old code used them), and the hosts parse the returned Stim DEM text with dem_from_stim_text. The runtime D_sparse now comes straight from the analysis' measurements-to-detectors
(m2d) map instead of generate_timelike_sparse_detector_matrix, remapped into each decoder's own enqueue stream for surface_code-3's split Z/X decoders (which run one dem_from_kernel pass per decoder via declare_detectors_z/_x
flags).

Notable constraint baked into the kernels:

• dem_from_kernel rejects kernels tagged qubitMeasurementFeedback, so the final data measurements are packed branch-free instead of via to_integer(to_bools(...)), whose discriminate-into-call flow sets the tag.

The now-unused compute_msm/construct_mz_table Python bindings (py_decoder.cpp, and their aliases in cudaq_qec/__init__.py) are removed. No other code references them after main's surface_code-1 refactor.

Note on the library API: main now offers decoder_context_from_memory_circuit(...) (with x_component()/z_component()/ full_component()), which surface_code-1 uses. surface_code-2/3 keep the hand-rolled dem_from_kernel path because that library circuit models noise only through the attached noise_model (gate channels), whereas these two examples deliberately demonstrate in-kernel SPAM (apply_noise data-qubit depolarization at p_spam); surface_code-3 additionally applies corrections as in-kernel transversal gates rather than the library's observable-flip model. Migrating them would change the modeled physics and measured logical error rate,
so it is intentionally out of scope here.

[eliotheinrich]

Looks good!

Removing the declare_detectors flags would be good, since the detector annotations should be automatically removed during compilation.

[melody-ren]

Looks good. Thanks! Just some notes about blurbs/stale looking comments

Should these two examples be referenced in the docs too? Currently they are not

[bmhowe23]

Re: referencing these in the docs — leaving them out for now; they're realtime-decoding test examples rather than user-facing tutorials.