Restoring lru (#896)

I did this before, but for some reason it wasn´t merged...

mathics/core/convert/mpmath.py

@@ -18,21 +18,7 @@
 from mathics.core.systemsymbols import SymbolIndeterminate
 
 
-# Another issue with the lru_cache: for mpmath, mpmath.mpc(0.,0.) and
-# mpmath.mpf(0.) produces the same hash. As a result, depending on
-# what is called first, the output of this function is different.
-#
-# note mmatera: When I start to pass the private doctests
-# to pytests, I realize that WMA evaluates `0. I`  to `Complex[0.,0.]`
-# instead of  `0.`. To fix this incompatibility, I removed from
-# `from_sympy` the lines that convert `mpc(0.,0.)` to `mpf(0.0)`,
-# and then this issue becomes evident.
-#
-# As we decide by now that performance comes after ensuring the compatibility
-# and clarity of the code, I propose here to conserve the right tests,
-# and commented out the cache and the lines that convert mpc(0,0) to MachineReal(0).
-#
-# @lru_cache(maxsize=1024)
+@lru_cache(maxsize=1024, typed=True)
 def from_mpmath(
     value: Union[mpmath.mpf, mpmath.mpc],
     precision: Optional[int] = None,
@@ -56,6 +42,13 @@ def from_mpmath(
         # HACK: use str here to prevent loss of precision
         return PrecisionReal(sympy.Float(str(value), precision=precision - 1))
     elif isinstance(value, mpmath.mpc):
+        # Comment mmatera:
+        # In Python, and mpmath, `0.j` and `0.` are equivalent, in the sense
+        # that are considered equal numbers, and have the same associated
+        # hash.
+        # In WMA, this is not the case. To produce the
+        # Python's behavior, uncomment the following lines:
+        #
         # if value.imag == 0.0:
         #   return from_mpmath(value.real, precision=precision)
         val_re, val_im = value.real, value.imag