Add Chebyshev approximations

[sigilante]

Prototype implementation demonstrating jet-compatible approach for transcendental functions (exp, sin, cos, tan) using:

• Fixed polynomial coefficients from musl libc
• Deterministic argument reduction
• Horner's method evaluation

Key principle: every FP operation happens in same order with same intermediate values in both Hoon and C jet.

[sigilante]

Follow-up analysis: libmath direction (from the Saloon eig work)

Captured here so it isn't lost — context is the Saloon eigendecomposition effort (PRs #47/#48), which surfaced concrete evidence for why this PR's approach is the right one. Pending for now; this is a roadmap, not a request to merge.

The problem this PR solves (with evidence)

The current math.hoon implements transcendentals as naive Taylor series with rtol-driven convergence loops (exp/sin/cos/log/sqt/… all loop until (lth (abs (sub po p)) rtol)). Three structural problems, worst last:

1. Unbounded → can hang the ship. While developing Hermitian eig (Saloon: Hermitian (%cplx) eigendecomposition (eig Phase A2) #48), a sub-ULP rtol made Newton sqt oscillate between two adjacent floats forever → loom exhaustion → recover: dig: meme + segfault (had to reboot with a bigger loom). This hazard exists in every rtol-loop transcendental, not just sqt.
2. Inaccurate. Argument reduction is minimal (sin does a single lossy (mod x tau); exp has none past overflow guards). Taylor converges slowly with cancellation — the doc comments show (cos pi) = -1.0000000000013558 (~1.4e-12 off, not f64-accurate). That's an accuracy ceiling for the whole Lagoon/Saloon stack.
3. Not bit-exactly jettable — the decisive one. Iteration count depends on value and rtol, and term order is series-specific, so a C jet can't reproduce the exact SoftFloat op sequence to yield an identical noun. Effectively unjettable as written.

Why the fixed-polynomial approach here is correct

This PR's stated principle — "every FP op happens in the same order with the same intermediate values in both Hoon and the C jet" — fixes all three at once:

• Fixed-degree minimax/Chebyshev polys with hex-exact frozen coefficients (musl is a fine reference) → bounded constant work → no hang.
• Deterministic argument reduction (Cody-Waite / Payne-Hanek) + Horner (or Clenshaw for a Chebyshev series) → near-correctly-rounded at low degree → accurate.
• Fixed straight-line op sequence → a C jet calling SoftFloat in the same order is bit-identical. Since vere's jets and Hoon @rd ops share the SoftFloat engine, bit-exactness is attainable — but only once the op order is fixed (which Taylor-with-variable-iterations is not).

Consequences / decisions for the migration

• rtol largely goes away. Fixed-degree functions take no tolerance; the [rnd rtol] door sample becomes vestigial for most arms. This also removes the exact foot-gun that crashed the ship (mismatched-width rtol).
• sqt is the natural first target and a clean template: Newton-for-sqrt with a fixed iteration count (5–6 steps after a good seed) is already deterministic and jettable — quadratic convergence makes it bit-stable. Saloon Saloon: Hermitian (%cplx) eigendecomposition (eig Phase A2) #48 ships an interim iteration-capped fsqt; the principled replacement is a jettable libmath sqt that Saloon then defers to.
• Clenshaw (Chebyshev series) vs Horner (minimax monomials): both are fixed-op-order and jettable. musl-style Horner with frozen coefficients matches the coefficient tables directly and is simplest for the jet path; the +cheb T_n recurrence in this PR is best viewed as derivation tooling, somewhat orthogonal to runtime evaluation. Recommend Horner for the jetted path, keep a Chebyshev generator as tooling.
• Payoff is automatic for Saloon/Lagoon: their element-wise exp/sin/… and the eig sqt route through libmath via trans-scalar, so converting libmath gives the array stack accurate + fast + jettable transcendentals for free.
• vs PR Add Chebyshev function support for more efficient calculations. #9 (sigilante/cheb): older and sprawls across lib3d/lagoon/math; likely superseded by this focused prototype. Salvage any useful coefficients, otherwise let it lapse.

Suggested first step (when picked up)

Convert sqt (all widths) to a fixed-iteration, rtol-free, bit-exact-jet-ready form as the migration template, then retire Saloon's interim fsqt. Small, self-contained, closes the durability hole, and establishes the pattern for exp/sin/cos/log.