Rename NoiseModel to NoiseParameters with a fluent interface

[ciaranra]

Summary

Renames the DEM-construction noise dataclass to NoiseParameters, lifts it to the
pecos top level, and gives it a fluent with_* interface in the shape of the
sim() noise builders. Closes #433.

Stacked on #420 — it depends on the shared idle-noise vocabulary introduced there.
Merge #420 first.

Why the rename

NoiseModel collided conceptually with the simulator-side noise models
(GeneralNoiseModel, DepolarizingNoiseModel), which are executable objects that apply
noise during simulation. This one is a parameter bag consumed at DEM construction, and
since #422 it is the shared idle-noise vocabulary for both the Guppy and native-surface
routes — so it was neither a "model" nor surface-code-specific.

• NoiseParameters is the real name; NoiseModel remains as a deprecated alias
  resolving to the identical class, so no existing code breaks.
• Exported from pecos top level, beside general_noise() / depolarizing_noise().
  Deliberately not placed in pecos.noise, which holds the legacy simulator error
  models (DepolarModel, XZModel, GatewiseModel); a modern DEM parameter class
  there would read as part of that old machinery.
• Still importable from pecos.qec.surface.
• 226 references updated across 29 existing files so only the alias path is deprecated,
  not everyday use.

Fluent interface

noise = (
    NoiseParameters()
    .with_p1(0.002)
    .with_p2(0.02)
    .with_p_idle_linear(0.01, {"X": 0.25, "Y": 0.25, "Z": 0.5})
    .with_p_idle_sin_squared(0.03, {"Z": 1.0})
)

Chained, immutable (every setter returns a new object via dataclasses.replace), and
with no terminal method — matching the sim noise builders, which are passed directly
into .noise(...). The dataclass constructor keeps working; the fluent form is additive.

One mechanical rule: with_<field_name>, no suffix translation. The simulator's
_probability / _rate suffixes were deliberately not mirrored: only four fields
(p1, p2, p_meas, p_prep) have simulator counterparts, so suffix-mirroring would
buy alignment on four methods while adding a translation layer to all thirty. It also
removes a hazard rather than documenting it — the simulator's with_p_idle_linear_rate
refers to what is our deprecated Z-only alias, not to the family scalar
p_idle_linear, so under the mechanical rule no false equivalence is expressible.

One deliberate exception, found by testing rather than by design: the idle families
take their rate and model together, as with_p_idle_linear(rate, model). Separate
setters are impossible in principle here — __post_init__ translates a family into the
canonical per-axis fields and then clears it, so a second call setting only the model
collides with the per-axis values the first call produced. A model without a rate is
inert and rejected anyway, so the two are inherently a pair. The first implementation
had split setters and raised ValueError on the natural usage
.with_p_idle_linear(0.01).with_p_idle_linear_model({"Z": 1.0}).

Verification

• tests/qec: 1393 passed, 1 skipped, 1 xfailed — this suite exercises most of the
  renamed call sites and is the oracle that behavior did not change.
• New tests/qec/surface/test_noise_parameters.py: 12 tests covering the equivalence
  oracle, immutability, the family round-trip through for_runtime_idle_time_units(),
  the deprecated alias, import paths, and the paired-setter rule.
• Repo-wide lint clean on a cold run with .ruff_cache cleared.