Fall back instead of baking a symbolic dim into softplus and nll_loss

[LeSingh1]

Same class as #2808: both converters read a shape dim and use it where only a constant works, so a flexible input shape aborts the conversion.

softplus with beta != 1 on a rank 4 input builds per-channel alpha/beta for softplus_parametric out of x.shape[1]:

model = torch.jit.trace(nn.Softplus(beta=2).eval(), torch.rand(2, 3, 4, 5))
ct.convert(model, inputs=[ct.TensorType(shape=(2, ct.RangeDim(2, 10), 4, 5))])
# TypeError: Cannot convert symbols to int

Falls back to the general decomposition softplus(beta * x) / beta when the channel dim is dynamic — the same path every other rank already takes, and numerically identical.

nll_loss with reduction="mean" divided by np.float32(x.shape[0]) before summing, which raises TypeError: Cannot convert expression to float on a dynamic batch. reduce_mean over the batch axis is the same value and needs no constant, so the batch size is no longer read at all. "none" and "sum" never used it and are unchanged.

Testing

test_softplus_beta_dynamic_channel and test_nllloss_dynamic_batch (all three reductions). 4 of the 8 cases fail on main — the "none" and "sum" reductions pass either way and are there as controls. Existing test_softplus and test_nllloss are unchanged and still pass. Checked nll_loss numerically at a batch size different from the traced one.

This machine has a broken scikit-learn install that makes every TorchFrontend.TORCHEXPORT case error out, including pre-existing ones, so I could only exercise the TorchScript frontend locally.