Bind sdr::OversampledChannelizer and OversampledSynthesizer

[Ravenwater]

Bind sw::dsp::sdr::OversampledChannelizer and OversampledSynthesizer — the
analysis/synthesis pair.

Upstream source: include/sw/dsp/sdr/channelizer.hpp (closes upstream #102).
Upstream tests: tests/test_sdr_channelizer.cpp.

Surface

• OversampledChannelizer<CoeffScalar, StateScalar, SampleScalar> —
  process(block) -> M complex channel outputs per hop, reset()
• OversampledSynthesizer<CoeffScalar, StateScalar, SampleScalar> — the
  inverse, M channels -> hop_ real samples, reset()
• Prototype-bank design helper (the h(N) / denom(hop) construction near
  the top of the header)

Binding notes

• This is not the same class as mpdsp.Channelizer. That one binds
  sw::dsp::multirate::Channelizer — critically sampled, hop = M. This one is
  oversampled: hop < M, so channels overlap and the analysis/synthesis pair
  reconstructs. Two classes with near-identical names in one namespace is a
  usability trap. Decide the Python names deliberately — OversampledChannelizer
  / OversampledSynthesizer mirrors upstream and is unambiguous; a bare
  Channelizer alias would collide with the existing binding.
• Also cross-link the two in the docstrings. Upstream's own docs added a
  pointer from multirate/channelizer.md to the synthesis half for exactly
  this reason (upstream 287584c).
• Three scalar axes — <CoeffScalar, StateScalar, SampleScalar>. This is
  the worst case for the dispatch decision in the foundations issue; at 18
  configs an independent-axis scheme is 5832 instantiations. If the foundations
  issue picked curated pairs, this class needs the triple form of that scheme.
• The prototype bank is designed once at construction. Like the RRC designer,
  CoeffScalar is design-time and narrows back to double for storage — so it
  may be expressible as coeff_dtype= at no runtime instantiation cost, which
  would reduce the live axes to two. Confirm against the header before
  assuming.
• Complex outputs go through complex_split_to_numpy; the 2-D (blocks, channels) shape follows the pattern already established in
  src/multirate_bindings.cpp::process_block.
• Note the upstream test lesson: a real input tone appears in two channels,
  the wanted one and its conjugate mirror at M - c. Upstream had to fix its
  own test for this (31a2ddb), and so did this repo's multirate channelizer
  tests. Write the Python test knowing it.

Depends on

Foundations issue (both decisions).

Deliverables

• Bindings in src/sdr_bindings.cpp.
• Exported by explicit name in python/mpdsp/__init__.py.
• tests/test_sdr_channelizer.py — tone lands in the expected channel and
  its conjugate mirror; analysis -> synthesis round trip reconstructs
  within tolerance; behaviour across the dtype table.
• scripts/build_api_ref.py tables updated, docs/api_reference.md
  regenerated.

---

Part of the SDR binding epic #130.