Validate public API arguments: chi_out=0 silently returns zero-size bonds #15

Problem

Neither apply() nor compress() validates its arguments, so invalid input
either produces silent garbage or surfaces as a confusing error from deep
inside NumPy or opt_einsum.

Measured behaviour on main:

The chi_out=0 case is the worst: a caller who computes chi_out from some
other quantity and gets 0 receives a well-formed but meaningless tensor
network with no indication that anything went wrong.

There is also an unchecked structural assumption: _, phys_up, phys_down = mpo.arrays[0].shape requires a rank-3 boundary tensor, so an MPO with
non-open boundary conditions fails with an opaque unpacking error.

Proposed fix

Validate at the public API boundary — in apply() and compress() — and raise
clear, actionable errors before any computation starts:

• chi_out must be an integer >= 1; raise ValueError otherwise, naming the
  argument and the offending value.
• cutoff must satisfy 0.0 <= cutoff < 1.0; it is a relative threshold on
  the largest singular value, so a value >= 1 discards everything and is
  never meaningful. Raise ValueError otherwise.
• In apply(), require left_tensor.nsites == right_tensor.nsites and raise a
  ValueError naming both counts.
• Check that the boundary tensors have the expected rank and raise a
  ValueError explaining that only open boundary conditions are supported.

Keep the messages specific enough to debug from a traceback alone, and follow
the existing EM/TRY ruff rules (message assigned to a variable first).

Acceptance criteria

• Each invalid input above raises ValueError with a message naming the argument and value.
• Validation happens before any array allocation.
• Parametrised tests cover every case in the table.
• Valid input behaviour is completely unchanged.