docs: better VI posterior and sampling methods how-to-guides

[janfb]

As discussed in #1765, the how-to guides around sampling methods and VI are a bit outdated and in parts misleading. This PR reworks them.

Restructure the sampler guide

09_sampler_interface was accurate but hard to act on: it opened abruptly and mixed
recommendations, configuration and low-level API across a flat list of sections. It now
follows the path a stuck user actually takes:

1. General context — why sampling differs between methods at all: NPE approximates the
   posterior directly so sampling is a forward pass, while NLE and NRE give a score
   proportional to the log posterior and need an extra algorithm.
2. Which sampler do you need? — routing by the inference method used, not by sampler.
   NPE/FMPE/NPSE: nothing to choose, plus the two exceptions (leakage, evaluable
   likelihood). NLE/NRE: start with MCMC, then branch to VI, rejection or importance
   sampling by the symptom you have.
3. Setting the sampler — build_posterior(sample_with=...) with PosteriorParameters.
4. Full control — the potential-function API as the low-level, high-flexibility
   interface, including building the sampler classes on a potential directly to use a
   non-prior proposal.

Give VI its own guide

VI had no guide of its own, so the sampler guide was the only place documenting
train() vs train_amortized(), the vi_method options and evaluate(). That content
moves to a new 26_variational_inference guide, which covers fixed-observation VI, quality
control with $\hat{k}$, SIR refinement, sequential NLE with VI, and amortized VI. The
sampler guide now links to it, as do the abstraction-levels and PosteriorParameters
guides and the implemented-methods tutorial.

Corrections

• Replace fixed parameter- and observation-count heuristics with guidance based on
  posterior geometry, proposal overlap, computational budget, repeated-observation needs
  and diagnostics.
• Fix the rejection-sampling acceptance-rate claim. RejectionPosterior.sample() discards
  the rate returned by rejection_sample(), so it cannot be read from the documented
  workflow; the sub-1% logging.warning is the only user-facing signal. Call
  sbi.samplers.rejection.rejection_sample() directly if the number itself is needed.
• Clarify the amortized ELBO/reverse-KL limitation and the mode-switching behaviour between
  train() and train_amortized().
• Scope importance sampling and SIR refinement to fixed-observation posteriors.
• Repair related inference-method links, including a link with a blank target.
• Add the vector-field guide to the how-to landing page; it was registered in the sampling
  toctree but missing from the list.

Every API name, keyword argument, default and behavioural claim in the new text was checked
against the source rather than carried over on trust.

Docs-only change, so no CHANGELOG entry and no tests. Built locally with Sphinx; all
cross-references resolve.

Claude Code assisted with the restructure and with verifying the API claims.

Closes #1765

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.66%. Comparing base (2c67b91) to head (eaa0c27).
⚠️ Report is 3 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1942      +/-   ##
==========================================
- Coverage   87.80%   87.66%   -0.14%     
==========================================
  Files         143      143              
  Lines       13435    14786    +1351     
==========================================
+ Hits        11796    12962    +1166     
- Misses       1639     1824     +185     

Flag	Coverage Δ
fast	81.61% <ø> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.
see 14 files with indirect coverage changes