Feature/PWPolyF

[ollycassidy13]

FINN needs a Brevitas-side PyTorch activation module that mirrors PWPolyF hardware behavior for common floating-point non-linear activations and can export a graph marker that FINN can lower to its RTL This PR adds a FINN-compatible piecewise-polynomial activation path for GELU, SiLU, sigmoid, and tanh, with QONNX export support through a finn.pwpolyf::PWPolyF custom-op marker.

The new activation:

• Supports gelu, silu, sigmoid, and tanh.
• Splits the FP32 input range into FINN/PWPolyF-style exponent and mantissa-derived segments.
• Fits configurable per-segment polynomial coefficients.
• Evaluates the approximation with Horner's method.
• Applies the expected clamp/pass-through behavior for |x| >= 8.
• Registers coefficients and clamp values as buffers so they are carried with the module.

[nickfraser]

Sorry, but this PR seems to be mostly vibe-coded and needs a LOT of work before it would be accepted. I will hold off from doing a full review before asking the following:

• do you really want to upstream this into Brevitas, rather than just using it as an out-of-source extension?
• if yes, are you willing to take the time to rearchitect this?

If the answer is 'no' to either of the above - please close this PR.

[nickfraser]

Remove this.

[nickfraser]

Remove this.