Fix pgm_read_float_unaligned macro (#6593)

Fixes #6590

The ASM block that implements the read-uint32-unaligned returns a
uint32_t.  The old code was doing a cast like `(float)(uint32_t ret)
which actually goes and creates a new float of the positive uint value
(approx, of course due to exponent and sign bits) which is not correct.

C and C++ don't have a concise way to convert the bits in a register
from int to float interpretation, so avoid the whole issue by making a
new function which uses the same ASM block as the read-uint32-unaligned,
just make the destination and return values as floats.

Author: earlephilhower

tools/sdk/libc/xtensa-lx106-elf/include/sys/pgmspace.h

@@ -80,6 +80,14 @@ static inline uint16_t pgm_read_word_inlined(const void* addr) {
   return (uint16_t) res;    /* This masks the lower half-word from the returned word */
 }
 
+/* Can't legally cast bits of uint32_t to a float w/o conversion or std::memcpy, which is inefficient. */
+/* The ASM block doesn't care the type, so just pass in what C thinks is a float and return in custom fcn. */
+static inline float pgm_read_float_unaligned(const void *addr) {
+  register float res;
+  pgm_read_with_offset(addr, res);
+  return res;
+}
+
 #define pgm_read_byte(addr)                pgm_read_byte_inlined(addr)
 #define pgm_read_word_aligned(addr)        pgm_read_word_inlined(addr)
 #ifdef __cplusplus
@@ -98,7 +106,6 @@ static inline uint32_t pgm_read_dword_unaligned(const void *addr) {
   return res;
 }
 
-#define pgm_read_float_unaligned(addr) ((float)pgm_read_dword_unaligned(addr))
 #define pgm_read_ptr_unaligned(addr)   ((void*)pgm_read_dword_unaligned(addr))
 #define pgm_read_word_unaligned(addr)  ((uint16_t)(pgm_read_dword_unaligned(addr) & 0xffff))