`KnownFPClass` and `APFloat` disagree on the definition of `subnormal` for `PPCDoubleDouble`.

The code for `KnownFPClass` is written with two assumptions about `fcSubnormal`:
1. `subnormal` values have a non-zero magnitude strictly less than `1.0`*.
2. `fcNegNormal < fcNegSubnormal < fcZero < fcPosSubnormal < fcPosNormal`.

*Or in practice, a value with a non-zero magnitude less than ~`0.01`, since the largest minimum normal value for a "common" floating point type is `2^-14 == 0.000061` for `ieee_binary16`.

However, `APFloat::classify` uses `APFloat::isDenormal` to determine if a value is `subnormal`. This is problematic for the `PPCDoubleDouble` type as values such as `{+1.0, DBL_TRUE_MIN}` are considered as `fcSubnormal` by `APFloat::classify`, breaking both of the common assumptions relied upon by `KnownFPClass`.

https://github.com/llvm/llvm-project/blob/0f391cb05e6e4727453b5ace37eda77ce67db85f/llvm/lib/Support/APFloat.cpp#L5716-L5721

https://github.com/llvm/llvm-project/blob/0f391cb05e6e4727453b5ace37eda77ce67db85f/llvm/lib/Support/APFloat.cpp#L5939-L5950

For `PPCDoubleDouble`, I was assuming that `fcSubnormal` was classified in one of two ways:
- The value is subnormal if the high part is subnormal.
- The value is subnormal if the value is non-zero and has a magnitude strictly less than `DBL_MIN * 2^53` (where it is unable to represent values that would require 106 bits of precision).

***

Possible solutions:
- Add `APFloat::isSubnormal`
- Add `fcDenormal`/`kfcDenormal` (As a follow up to the sequence of PRs outlined by https://github.com/llvm/llvm-project/issues/217072).
- Pass the floating point semantics into each `KnownFPClass` function, and return `fcAllFlags`/unknown if we have not tested or implemented the function for a given floating point semantics (could also be done via virtual functions or templates). This would allow `KnownFPClass` to only support the types we have properly verified, instead of excluding the types we have found bugs for.

As a follow up, it may be good to consider if non-canonical x87 forms should have dedicated classification flags (such as unnormal numbers, pseudo-subnormal, pseudo-NaN, etc)

Start with the linked sections of llvm/lib/Support/APFloat.cpp, especially APFloat::classify and isDenormal, then trace the KnownFPClass assumptions about fcSubnormal. Determine and document a consistent classification for PPCDoubleDouble, with the chosen solution resolving the disagreement without breaking the stated ordering or magnitude assumptions.

Written by the indexing model from the issue text.