Bind end-to-end SDR precision analysis (analysis/sdr_precision.hpp)

[Ravenwater]

Bind the end-to-end SDR precision analysis — the per-block precision sweep
that is the point of the whole module.

Upstream source: include/sw/dsp/analysis/sdr_precision.hpp (closes upstream
#103). Upstream tests: tests/test_sdr_precision.cpp.

Surface

• enum class SdrBlock { none, constellation, tx_shaping, rx_matched, whole_chain } + to_string
• SdrLinkConfig — modulation, samples_per_symbol, rolloff, rrc_span,
  ebn0_db, num_symbols, seed
• SdrLinkResult — block, scalar_type, bit_width, modulation,
  ebn0_db, evm_rms, and the rest of the measured fields
• run_link<Narrow>(cfg, block) -> SdrLinkResult
• analyze_blocks<Narrow>(cfg, ...) -> std::vector<SdrLinkResult>
• SdrParetoPoint
• evm_budget(Modulation) -> double

Binding notes

• This is the direct analogue of the acquisition precision primitives already
  bound in src/analysis_bindings.cpp (AcquisitionPrecisionRow,
  CICBitGrowthReport, write_acquisition_csv). Follow that file's shape,
  including a write_sdr_csv counterpart if upstream has one — check before
  inventing one.
• run_link / analyze_blocks are templated on Narrow and dispatch through
  dispatch_dtype_fn. Single axis — dtype= is correct here, no scalar-pair
  scheme needed. The link is double throughout; Narrow selects which named
  block gets projected, which is what SdrBlock enumerates.
• SdrLinkConfig.seed is unsigned and the link is stochastic. Bind the seed
  and default it, so Python results are reproducible; tests must set it
  explicitly.
• analyze_blocks over all five SdrBlock values across the dtype table is
  the expensive call. Note the runtime in the docstring.
• Carry upstream's corrected premise into the docstrings. Upstream issue
  0.6.0 Phase 2: expose CoeffScalar/T in window and FIR design bindings #85's epic assumed posit reaches a higher modulation order than cfloat or
  fixpnt at equal bit width; the bindings gap Phase 3: bind RealtimeSpectrum + detector modes #104 sweep measured the opposite at full
  scale, because an amplitude-normalized link keeps the waveform inside one
  octave and a uniform step is optimal against an absolute error metric there.
  If the Python docstrings restate the old premise they will actively mislead.
  Read upstream's docs-site/src/content/docs/sdr/precision.md first.
• Result structs bind with def_ro scalar properties.

Depends on

Constellation issue (Modulation enum); metrics issue (EvmResult shape
informs SdrLinkResult). Does not depend on the scalar-pair decision.

Deliverables

• Bindings — decide whether these live in src/sdr_bindings.cpp or
  src/analysis_bindings.cpp. The upstream header is under analysis/,
  which argues for the latter and keeps the precision primitives together.
• Exported by explicit name in python/mpdsp/__init__.py.
• tests/test_sdr_precision.py — a fixed seed reproduces exactly;
  SdrBlock::none is the double reference; narrowing whole_chain
  degrades EVM monotonically relative to none.
• scripts/build_api_ref.py tables updated, docs/api_reference.md
  regenerated.

---

Part of the SDR binding epic #130.