asymmetric_loss_channel (tableau) ignores measurement back-action #249

Where: crates/ppvm-tableau/src/noise.rs, GeneralizedTableau::asymmetric_loss_channel

Problem
The total loss probability is computed correctly from the populations,
p_tot = p0·P(0) + p1·P(1), but the state update after sampling is not the one
given by the Kraus operators

K0 = √(1−p0)|0⟩⟨0| + √(1−p1)|1⟩⟨1| + |L⟩⟨L|,   K1 = √p0 |L⟩⟨0|,   K2 = √p1 |L⟩⟨1|

For p0 ≠ p1 the channel acts as a weak Z measurement, but the implementation
doesn't account for this:

• Lost branch: the qubit is collapsed with an ordinary measure (Born
  probabilities P(0), P(1)) before reset. It should collapse to |0⟩ with
  probability p0·P(0)/p_tot and to |1⟩ with p1·P(1)/p_tot.
• Survival branch: the state is left unchanged. It should be updated with
  K0 (renormalized), which biases the survivor towards the less lossy state.

For an isolated qubit this is invisible (the lost qubit is reset anyway), but
correlations with other qubits come out wrong.

Example
Bell pair (|00⟩ + |11⟩)/√2, asymmetric loss on qubit 0 with p0 = 0, p1 = 1:

Possible fix
Sample one of three branches: lost-from-|0⟩ (p0·P(0)), lost-from-|1⟩ (p1·P(1)),
or survived (1 − p_tot).

• Lost branches: project onto the corresponding Z outcome (forced-outcome
  measurement), then reset and flag as lost.
• Survived: apply K0 restricted to the computational subspace,
  a·I + b·Z with a = (√(1−p0)+√(1−p1))/2 and b = (√(1−p0)−√(1−p1))/2, then renormalize.
  This is a non-Clifford (non-unitary) Pauli-sum update, which fits the
  generalized tableau's coefficient vector. It is trivial if the qubit is in a Z eigenstate.

Scope
Only reachable via the Rust/Python API; the bytecode ISA has no asymmetric
loss instruction. Symmetric (loss_channel) and correlated loss are unaffected,
since there the Born-rule collapse coincides with the conditional distribution.
leakage_channel takes p0/p1 as pin probabilities rather than state-dependent
loss rates, so it isn't affected either.