Validate calculate_ims waveform shape and support (nt, 3) input

[lispandfound]

Closes #231

What was wrong

calculate_ims (IM/im_calculation.py) used np.atleast_3d followed by
np.moveaxis(-1, 0) to reshape the input waveform, but the docstring only
said "Waveform data as a NumPy array" and never stated that the only layout
this actually supports is a single station shaped (1, nt, 3):

• A more intuitive (nt, 3) array was silently reshaped by atleast_3d into
  (nt, 3, 1), which then became (1, nt, 3) after moveaxis — a
  single-component waveform — and failed deep inside IM/ims.py with a
  confusing TypeError: Waveform must have 3 components, but waveform.sizes['component']=1.
• A multi-station array (n, nt, 3) with n > 1 passed the component check
  but blew up later in _dataset_to_frame with a pandas index length
  mismatch, since that function assumes a single station's worth of rows.

Fix

• Explicitly validate waveform's shape instead of relying on
  np.atleast_3d: accept (nt, 3) (by adding the leading station axis) or
  (1, nt, 3), and raise a clear ValueError for anything else, including
  multiple stations.
• Updated the docstring to state the accepted shapes and component order
  (000, 090, ver).

Test

Added two tests to tests/test_ims.py:

• test_calculate_ims_accepts_2d_waveform: passes both a (1, nt, 3) and the
  equivalent (nt, 3) waveform through calculate_ims and checks the results
  match. Before the fix, this failed with the TypeError about component
  count quoted in the issue.
• test_calculate_ims_rejects_multiple_stations: passes a (2, nt, 3)
  waveform and checks calculate_ims now raises a clear ValueError. Before
  the fix, this failed with a pandas "Length mismatch" error instead.

Both tests were confirmed to fail against the pre-fix code (for the reasons
above) before the fix was applied, and pass afterwards. .ai/ci-checks was
run in full and passes.

What to check closely

• The new shape validation only accepts waveform.shape == (1, nt, 3) or
  (nt, 3); if there is a use case elsewhere in the codebase that relies on
  calculate_ims accepting multiple stations at once, that call site would
  now need to loop over stations instead (a search of the repo didn't turn up
  any such caller).

Cost $0.69, 43 turns.

🤖 Generated with Claude Code

[lispandfound]

@joelridden please look at this closely and check it isn't actually a documentation bug instead of a code issue.

[joelridden]

All this is really doing is making sure if we use another function other than waveform_reading.read_ascii (Which already does the reshaped_waveform = waveform_data[np.newaxis, :, :]) to read a waveform and pass it in, to do the same step. Since NZGMDB is using the read_ascii function anyway this is fine to do, I don't consider it that "needed" though, but could be nice to have in future if there is other sources of waveforms to be read and passed into this IM calc function.