[read-fonts] Speed up ValueFormat size calculation

[behdad]

I see 0.8% in my Roboto HarfRust profile spent in value_record.rs:

    /// Return the number of bytes required to store a [`ValueRecord`] in this format.
    #[inline]
    pub fn record_byte_len(self) -> usize {
        self.bits().count_ones() as usize * u16::RAW_BYTE_LEN
    }

As it happens, we're only interested in counting ones of a u8 number. In HB I optimized this by writing a custom function instead of relying on the compiler intrinsic:

/* Return the number of 1 bits in a uint8_t; faster than hb_popcount() */
static inline unsigned
hb_popcount8 (uint8_t v)
{
  static const uint8_t popcount4[16] = {
    0, 1, 1, 2, 1, 2, 2, 3,
    1, 2, 2, 3, 2, 3, 3, 4
  };
  return popcount4[v & 0xF] + popcount4[v >> 4];
}

This is the case because unless you build for your native arch, compilers don't rely on CPU ops for popcount and generate code that takes about 10 ops (branchless). I found that this replacement speeds up measurably and reliably.

Caveat:

  // Note: spec says skip 2 bytes per bit in the valueformat. But reports
  // from Microsoft developers indicate that only the fields that are
  // currently defined are counted. We don't expect any new fields to
  // be added to ValueFormat. As such, we use the faster hb_popcount8
  // that only processes the lowest 8 bits.