The issue #483 is solved with the following logic:
1. Fix TorchBaseOptimizer and parameter space consistency

Files: base.py in torch optimizers

torch optimizers now work consistently in scaled parameter space, matching the scipy optimizers:

• Init: var.variable.get_value() → var.value (scaled)
• Update: var.variable.update_value(param) → var.update(param) (inverse-scales internally)
• Bounds: var.bounds (unchanged — already scaled)

This ensures _apply_bounds() clamps parameters in the same space they live in, preventing catastrophic value corruption.

The issue #484 is solved with the following logic:

2. Material n / k caching causes RuntimeError under torch + grad mode

When using the torch backend with gradient tracking enabled, calling .backward() more than once during optimization raises:
RuntimeError: Trying to backward through the graph a second time
(or directly access saved tensors after they have already been freed).

Why this happens: When BaseMaterial.n() computes a refractive index under torch, the result is a tensor connected to a computation graph. If that tensor is cached and reused in a later forward pass, the new .backward() call tries to traverse the old (already freed) graph — causing the RuntimeError.

The previous fix I applied privately bypassed the cache entirely whenever be.grad_mode.requires_grad is True:

if be.get_backend() == "torch" and be.grad_mode.requires_grad:
    return self._calculate_n(wavelength, **kwargs) # skip cache'

This solved the RuntimeError but introduced two problems:

Performance: Every call recomputes the refractive index from scratch, even for identical wavelengths. During optimization, the same materials are queried thousands of times — a significant penalty.
Broke the caching test: test_caching[backend=torch] failed because the cache was never populated.

The fix in this PR user a more robust approach - before caching, check what kind of result we got:

• If the result requires grad (i.e., the index itself is an optimization variable like IdealMaterial(n=nn.Parameter(...))): skip cache, return fresh — gradient flow must be preserved.

• If the result does not require grad (i.e., it's a constant like Material("N-BK7")): detach and cache — .detach() severs the link to the old computation graph so it can't cause the RuntimeError, and caching avoids redundant recomputation.

This gives us correct gradient behavior, no RuntimeError, and full caching performance for the common case where materials are constants.

Fixes #483 and #484