[enhancement] Invariant assertion for one-hot selector should codegen with the RTL net reference, not the IR node.

### What's hard to do? (limit 100 words)

It is confusing for the generated assertion to reference the XLS IR node name in the RTL assertion. E.g. from https://github.com/google/xls/issues/3288, we get:

```
`ifdef ASSERT_ON
__xls_invariant_x26_squeezed_selector_one_hot_A: assert property (@(posedge clk) disable iff ($sampled(rst !== 1'h0 || $isunknown(p2_and_2553_comb))) p2_and_2553_comb) else $fatal(0, "Selector concat.124 was expected to be one-hot, and is not.");
`endif // ASSERT_ON
```

But `concat.124` is not found in the RTL, only in the .opt.ir file. This can be particularly surprising to hardware engineers whose first reaction is to debug the RTL.

### Current best alternative workaround (limit 100 words)

Refer back to the opt IR file.

### Your view of the "best case XLS enhancement" (limit 100 words)

@grebe made the following comment in chat:

> Yeah, this happens because we bake the string when the assert is first made. You could imagine something like how trace does format strings with a format specifier for "name of arg". My guess is it wouldn't be too much work. It might make the assert more of an optimization barrier?