fix: dataset guess

[100pah]

(1) Enhance dataset default encode guess strategy. Fix #11428, Fix #11419, Fix #11735
(2) Support candlestick in dataset encode guess.
(3) Some refactor of the code arrangement of dataset encode guess and dimension calculation.

Break change of:
<test/dataset-charts.html>

The test case:
<test/dataset-guess.html>

Strategy:

line, bar, pictorialBar, scatter, effectScatter, candlestick, boxplot

[The strategy of the arrengment of data dimensions for dataset]:

• "value way": if all axes are non-category axes. So series one by one take
  several (the number is coordSysDims.length) dimensions from dataset.
  The result of data arrengment of data dimensions like:
  | ser0_x | ser0_y | ser1_x | ser1_y | ser2_x | ser2_y |
• "category way": if at least one axis is category axis. So the the first data
  dimension is always mapped to the first category axis and shared by
  all of the series. The other data dimensions are taken by series like
  "value way" does.
  The result of data arrengment of data dimensions like:
  | ser_shared_x | ser0_y | ser1_y | ser2_y |

pie, funnel, map

Find the first pure number dim as the value dim,
and then find a name dim with the priority:
"number-like|other string dim" > "other dim" > "the value dim itself".

If the previous strategy is not satisfied (that is, no pure number),
find the first number-like dim as the value dim,
and then find a name dim with the priority:
"other dim" > "the value dim itself".
That is for backward compat: number-like (e.g., '3', '55') can be
treated as number.