Support approximately-equal floating point assertions using ULPs

Google's C++ test framework supports [approximate comparison of floating point numbers ](https://github.com/google/googletest/blob/master/googletest/docs/advanced.md#floating-point-comparison) without requiring the user specify a precision (though that's available too). The `ASSERT_DOUBLE_EQ`/`ASSERT_FLOAT_EQ` macros assert that two numbers are [within four ULPs of one another](https://github.com/google/googletest/blob/472cd8fd8b1c665bddfd021ad0f62d6747fe8e72/googletest/include/gtest/internal/gtest-internal.h#L346), which (the documentations claims) should work well enough for most use-cases.

I'm *sure* this has been discussed before, so apologies if I'm dredging up something that's already been long-decided. But if ULP-based assertions are intentionally absent from `DoubleSubject` and `FloatSubject` I think the documentation should explain why that decision was made and what users should do instead. Currently the class makes no mention of ULPs, and expects callers to always pick an appropriate `tolerance`.