Provide proof that `tan(x)` does not overflow to infinity for `fp80`, `fp128`, and `ppcf128`.

Currently, no proof is provided that `tan(x)` never overflows to infinity in `KnownFPClass::tan`.
https://github.com/llvm/llvm-project/blob/25525299adf55ea46d8590ef4385ea66f2679822/llvm/lib/Support/KnownFPClass.cpp#L664-L669

You can brute force prove that a correctly rounded `tan(float)` will never overflow to infinity, and it seems like `tan(double)` will never overflow either https://stackoverflow.com/questions/67482420/proof-of-impossibility-or-counterexample-for-tanx-infinity-for-floating-p

However, it seems like there is no proof that `tan(x)` won't overflow for `f80`, `f128`, or `ppcf128`. Also, since `f80` represents a subset of the values from `f128`, a proof that `tan(f128)` never overflows automatically proves that `tan(f80)` will also never overflow. If I were to guess, `ppcf128` is probably the most likely to overflow since it has the smallest exponent.

I suspect the existing deduction is correct, but brute-force testing cannot rule out an exceptionally close representable approximation to a pole for `tan(x)`.

***

Using this AI assisted script, I brute force checked all poles up to 100'000'000.5 * pi for `ppcf128` at 160 bits of precision. The script would compute `tan(x)` where `x` is `(pole_number + 0.5) * pi` rounded to the nearest `ppcf128` value.
https://pastebin.com/KeaqrwTq

This result was the closest to overflowing that I could find (required 112 bits of precision to detect, and I verified the result at 2112 bits of precision)
```
x = 5564510.5 * pi
x = 0x1.0abed14ec066p+24 + -0x1.83ec193dafea8p-38
tan(x) = 0xb.44d4b3884399665e566777e455a18e8e2ecf841bp+108
```

The target is llvm/lib/Support/KnownFPClass.cpp, especially KnownFPClass::tan at lines 664-669. Start by tracing the existing no-overflow deduction and comparing the representable ranges of fp80, fp128, and ppcf128, using the issue's pole-search result as context. Done means the deduction is backed by a valid proof for all three formats, or is no longer asserted where it remains unproven.

Written by the indexing model from the issue text.