Add functional transcendentals + scalar-affine ops for scientific forward models

[NicolasRouquette]

A small, single-purpose contribution: surface differentiable elementwise exp /
log and scalar-affine maps on the nn.functional API, so scientific forward
models — radiative transfer, dielectric mixing, kinetics, anything built from an
exp/log of an affine argument rather than from NN activations — can be written
as a pure autograd.fn1.Fn and differentiated by the autograd engine, with no
hand-coded gradient.

Context

Independent and self-contained: this branch is based directly on main and does
not depend on the factorization PRs (#2 and its eig/SVD follow-up). The diff is 7
files, +254 lines.

The motivating use case is a soil-moisture retrieval that combines SMAP (Soil
Moisture Active Passive) and NISAR (NASA–ISRO Synthetic Aperture Radar)
observations through the AVS (Attenuation–Volume–Surface) model, whose surface
term is exp(-2·b·NDVI)·c·|R|². Operationally that model is fit per pixel with a
hand-coded analytic Jacobian (plus a byte-duplicated JIT copy). A sign or factor
error there does not crash anything — it silently degrades the fit, and the copies
can drift. The ops here let the forward model be written once and its gradient be
derived by autograd, making the hand-coded Jacobian redundant.

What's here

• Functional ops (NN/Runtime/Autograd/TorchLean/Functional/Core.lean):
  F.{exp, log, scale, shift, affine} — thin lifts of the existing
  Ops.{exp, log, scale, const} primitives, which already carry registered
  backwards, so reverse-mode jacrev / grad works through them unchanged.
  affine x c k = c·x + k; scale / shift are the multiply / add-by-constant
  halves.
• Public surface (NN/API/Public/NN/FunctionalBatch.lean,
  NN/API/Public/Seeded.lean): the five ops added to both nn.functional export
  bridges, beside square / mean.
• Example + self-check (NN/Examples/Functional/Transcendentals.lean,
  lake exe transcendentals_check): differentiates exp, 3x+1, and exp(-2x)
  and compares the autograd gradient to the closed form. Positive controls plus a
  negative control — the autograd gradient of exp(-2x) is -2·e^{-2x} and the
  test rejects the wrong-sign +2·e^{-2x}, i.e. exactly the hand-coded-Jacobian
  defect class the ops are meant to eliminate. Exits non-zero on regression.
• Blueprint (blueprint/.../Guide/Ch2_Frontend/ScientificForwardModels.lean):
  a short chapter documenting the ops and the autograd-derived-gradient point.

Notes for reviewers

• The example links the toolchain libc++ at runtime (native autograd backend);
  run with LD_LIBRARY_PATH=<toolchain>/lib.
• Nothing in Ops or the backends changes; this is purely additive on the
  functional surface.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com