ENH: use minimax approximations for real branches of lambertw (#119)

* initial fukushima implementation

* create dld_d

* keep tol for compatibility

* fix polynomial coefficient ordering

* add k=-1 branch

* create flf_f

* add float overload

* fix minus sign

* fix polynomial orders

* fix W-1 coefficient tables mixup

* undo namespace split; use cephes::ratevl

* minor cleanups

* copy tests from #116

* fix tests

* more test tweaks

* nan/inf tweaks

* run `pixi run format`

* undo changes to evalpoly.h

* Apply suggestions from code review

Co-authored-by: Albert Steppi <1953382+steppi@users.noreply.github.com>

* define constants inside the function body

* move coeff arrays to nested detail::lambertw_real namespace

* add special-case shortcut to complex overload

* pixi run format

* address review comments

* fix suspected coefficient typo

* experimental series expansion around zero

* bugfix

* improve series expansion

* cleanup

* tweak test tolerance following review

---------

Co-authored-by: Albert Steppi <1953382+steppi@users.noreply.github.com>

Author: kandersolar

include/xsf/cephes/polevl.h

@@ -116,7 +116,7 @@ namespace cephes {
 
     /* Evaluate a rational function. See [1]. */
 
-    /* The function ratevl is only used once in cephes/lanczos.h. */
+    /* The function ratevl is only used in lambertw.h and cephes/lanczos.h.*/
     XSF_HOST_DEVICE inline double ratevl(double x, const double num[], int M, const double denom[], int N) {
         int i, dir;
         double y, num_ans, denom_ans;