fix -m 2 mode-cost evaluation (causing partition0 overflow)

The mode's bits were not taken into account, which is ok for most of cases.
But in case of super large image, with 'easy' content, their overhead starts
mattering a lot and we were omitting to optimize for these.
Now, these mode bits have their own lambda values associated, limiting
the jerkiness. We also limit (for -m 2 only) the individual number of bits
to something that will prevent the partition 0 overflow.

removed the I4_PENALTY constant, which was a rather crude approximation.
Replaced by some q-dependent expression.

fixes issue #289

Change-Id: I956ae2d2308c339adc4706d52722f0bb61ccf18c

src/enc/webpenc.c

@@ -105,6 +105,10 @@ static void MapConfigToTools(VP8Encoder* const enc) {
       256 * 16 * 16 *                 // upper bound: up to 16bit per 4x4 block
       (limit * limit) / (100 * 100);  // ... modulated with a quadratic curve.
 
+  // partition0 = 512k max.
+  enc->mb_header_limit_ =
+      (score_t)256 * 510 * 8 * 1024 / (enc->mb_w_ * enc->mb_h_);
+
   enc->thread_level_ = config->thread_level;
 
   enc->do_search_ = (config->target_size > 0 || config->target_PSNR > 0);