Fix compile errors with msvc9 and msvc10

[cgohlke]

I took the definition of NAN from Cython.
See also #128

[jennolsen84]

@kwgoodman what if we changed line 14 to:

const double NAN = __NAN()
const double NAN64 = __NAN()
const float NAN32 = __NAN()

This will avoid calling memset() repeatedly, as there will be only one NAN value. Also, later down the line, we can turn move_median into a template so that it is fast for float32 as well.

[kwgoodman]

@jennolsen84, I had trouble redefining NAN using const double NAN = __NAN(). I ended up trying:

kg@kg:/devel/bottleneck (cgohlke-patch-1 *)$ di
diff --git a/bottleneck/csrc/move_median.c b/bottleneck/csrc/move_median.c
index cf94d83..ff5d477 100644
--- a/bottleneck/csrc/move_median.c
+++ b/bottleneck/csrc/move_median.c
@@ -10,6 +10,11 @@ node1->idx = idx2;                                 \
 node2->idx = idx1;                                 \
 idx1       = idx2

+/* define nan */
+ai_t myNAN;
+void init_nan(void) {
+  memset(&myNAN, 0xFF, sizeof(ai_t));
+}

 /*
 -----------------------------------------------------------------------------
@@ -64,6 +69,8 @@ mm_new(const idx_t window, idx_t min_count)

     mm_reset(mm);

+    init_nan();
+
     return mm;
 }

@@ -511,7 +518,7 @@ mm_get_median(mm_handle *mm)
 {
     idx_t n_total = mm->n_l + mm->n_s;
     if (n_total < mm->min_count)
-        return NAN;
+        return myNAN;
     if (min(mm->window, n_total) % 2 == 1)
         return mm->s_heap[0]->ai;
     return (mm->s_heap[0]->ai + mm->l_heap[0]->ai) / 2.0;
diff --git a/bottleneck/csrc/move_median.h b/bottleneck/csrc/move_median.h
index d6bbe54..736b0b0 100644
--- a/bottleneck/csrc/move_median.h
+++ b/bottleneck/csrc/move_median.h
@@ -6,12 +6,6 @@

 #if defined(_MSC_VER) && (_MSC_VER < 1900)
 #define inline __inline
-static __inline float __NAN() {
-  float value;
-  memset(&value, 0xFF, sizeof(value));
-  return value;
-}
-#define NAN __NAN()
 #endif

 #if NOCYTHON==1

Note that NAN is only used in one place---calculating the median when the count is less than min_count. These changes slow the code a bit, but I wasn't too careful in my measurement.