Normalize eig_trunc_pullback! iteration by the smallest eigenvalue and use _smith_iteration!

[leburgel]

Stacked on #287, which introduces a dedicated helper for the doubling-iteration Sylvester solve.

eig_trunc_pullback! still normalized its Sylvester iteration by the largest |eigenvalue|, so max|D̄⁻¹| = max|D| / min|D| can grow under repeated squaring and overflow when the retained eigenvalues span a large range. This is the same issue #282 fixed for eigh_trunc_pullback!.

Changes:

• Normalize by the smallest |eigenvalue| instead, which caps max|D̄⁻¹| at 1 so squaring can only shrink it.
• Replace the inline doubling loop with the shared _smith_iteration! helper, as already done for eigh_trunc_pullback! and svd_trunc_pullback! in Sum only one Neumann series in svd_trunc_pullback! #287.
• Build APᴴ = A' - ViG * (V * D)' directly rather than AP, so the helper receives a plain matrix instead of an Adjoint, which avoids generic fallbacks on GPU.

Behavioural note: as for eigh_trunc_pullback! in #287, the first term APᴴ X₀ D̄⁻¹ is now also subject to the degeneracy_atol stopping test instead of always being added.