Create wmin/wmax on the device of the given weights - #837

What this changes and why

A connection built from weights already on a GPU (Connection(..., w=w_cuda, wmin=0.0, wmax=1.0))
failed in its constructor with Expected all tensors to be on the same device, but got min is on cpu.
AbstractConnection.__init__ always created wmin and wmax on the CPU, and the subclass
constructors clamp the given w against them. The bounds are now created on the device of the
given w. Without w the weights start on the CPU, and so do the bounds, as before.
Reported on Intel XPU (Aurora); the same error happens on CUDA.

Casting the bounds to float avoids the error only for single-value bounds; per-weight bounds
(a tensor the size of w, which the docstring allows) cannot be cast. This fix covers both.

Checklist

• I ran poetry run pytest locally and it passed. (644 passed, on a machine with CUDA)
• If I fixed a bug, this test fails on master and passes with my change
  (checked by the bug-test check):
  test/network/test_connections.py::test_connection_from_gpu_weights_clamps_and_learns_on_gpu
  It needs CUDA, so on CI (no GPU) it is skipped and bug-test cannot see it. On a CUDA
  machine both parameter cases fail on master with the device error and pass with the fix.
• I did not remove tests, remove checks, add skips or xfails, or widen tolerances.
  The new GPU test carries skipif(not torch.cuda.is_available()), like
  test_windows_and_maxpool_state_stay_on_gpu in the same file.
• If I changed a formula: not applicable.
• Does this change simulated results (spikes, voltages, traces, weights)?
  - [x] No. Only the device of the bounds changes, and only when w is given.
• Did any of these files change? No.
• I added an entry to the Unreleased section of CHANGELOG.md.

AI assistance

• No AI tools were used.
• AI tools were used, and I (a human) have read every changed line, run the tests myself, and can explain the change in my own words.

Tools used and for which parts: Claude Code reproduced the bug, wrote the fix, the tests and the
changelog entry.

Not in this pull request

Conv1dConnection, Conv2dConnection, Conv3dConnection and MeanFieldConnection clamp a given
w only when a bound is infinite ((self.wmin == -inf).any() or (self.wmax == inf).any()), the
reverse of Connection (!=). With finite bounds a given w is never clamped. Fixing it changes
results for anyone passing out-of-range weights, so it belongs in its own pull request.