feat(fluo): FISTA (non-negative) reconstruction for 3D fluorescence

[talonchandler]

Summary

Adds reconstruction_algorithm: FISTA to 3D fluorescence reconstructions. FISTA solves

```
min 0.5 ||Hx - y||^2 + 0.5 lambda ||x||^2 s.t. x >= 0
```

via accelerated proximal gradient with a ReLU prox at every iteration, guaranteeing a non-negative reconstruction. Shares regularization_strength with the existing Tikhonov path.

Config example

```yaml
reconstruction_dimension: 3
input_channel_names: [GFP]
fluorescence:
transfer_function:
yx_pixel_size: 0.1
z_pixel_size: 0.25
wavelength_emission: 0.532
index_of_refraction_media: 1.3
numerical_aperture_detection: 1.2
apply_inverse:
reconstruction_algorithm: FISTA # NEW
regularization_strength: 5.0e-3 # shared L2 with Tikhonov
FISTA_max_iter: 100 # NEW
FISTA_rel_change_tol: 1.0e-3 # NEW
```

Full file at benchmarks/configs/fluor_3d_beads_fista.yml.

Motivation

Standard TLS / Tikhonov fluorescence reconstructions produce non-physical negative voxels (~50% in dynacell mantis A549 SEC61 ER and hummingbird A549 G3BP1 stress granule data) regardless of regularization choice. FISTA removes the negatives by construction at the cost of ~30-100x more wall time (still well under a minute per 100-iter, 512^2 lateral FOV on CPU).

Scope

• Only 3D fluorescence (recon_dim=3). 2D fluorescence (singular-system path) explicitly raises NotImplementedError for FISTA for now.
• Phase / birefringence settings untouched. The FISTA option lives on a new ApplyInverseSettings subclass in waveorder/api/fluorescence.py rather than the shared FourierApplyInverseSettings, so phase / birefringence schemas stay clean.

Changes

• waveorder/api/fluorescence.py: new ApplyInverseSettings(FourierApplyInverseSettings) widens the algorithm Literal and adds FISTA_max_iter / FISTA_rel_change_tol. Apply-inverse dispatcher filters FISTA-only kwargs before calling the 2D entry-point and rejects 2D FISTA with a clear message.
• waveorder/models/isotropic_fluorescent_thick_3d.py: FISTA branch with constant step size 1 / L, L = max(|OTF|^2) + lambda. Two FFTs per iteration (FFT(z), IFFT(residual_fft * conj(OTF))), early stop on ||x_{k+1} - x_k|| / ||x_k|| < FISTA_rel_change_tol.
• benchmarks/configs/fluor_3d_beads_fista.yml: example config.

Tested

Smoke test on simulate(...) phantom in 3D: Tikhonov and FISTA both reconstruct the bead within a few counts of the true intensity; FISTA stays strictly non-negative. 2D FISTA correctly raises NotImplementedError.

```
phantom range: 0.0 1.0
data range: 10.07 10.55
Tikhonov out: min=9.57 max=10.90 frac<0=0.000
FISTA out: min=9.49 max=10.80 frac<0=0.000
OK: FISTA in 2D raised: FISTA reconstruction is currently only implemented for 3D fluorescence
```

Follow-ups (not in this PR)

• 2D FISTA in isotropic_fluorescent_thin_3d.apply_inverse_transfer_function (different forward model — singular-system convolution).
• Per-voxel Poisson-with-offset model for low-SNR data.

Test plan

• Tikhonov path produces same results as before (regression).
• FISTA on a simulated 3D phantom converges and stays non-negative.
• 2D FISTA raises NotImplementedError.