916b8e let code executed with mrb->exc set, and may cause a crash
like #3519. Instead modified OP_RESCUE again. To retrieve the exception
object, we use `OP_RESCUE R(A), 0, 0` (old behavior). To compare the
exception object and the class, we use `OP_RESCUE R(A), R(B), 1`.

The reason we use OP_RESCUE for two instruction switched by operand
C is to save the instruction space.

As a result, the following code:
```ruby
begin
  raise "a"
rescue TypeError
  p 1
rescue RuntimeError
  p 2
end
```
will be compiled as:
```
irep 0x55cd1f565cb0 nregs=4 nlocals=1 pools=1 syms=4 reps=0
file: -
    2 000 OP_ONERR	005
    2 001 OP_LOADSELF	R1
    2 002 OP_STRING	R2	L(0)	; "a"
    2 003 OP_SEND	R1	:raise	1
    2 004 OP_JMP	023
    2 005 OP_RESCUE	R1
    3 006 OP_GETCONST	R2	:TypeError
    3 007 OP_RESCUE	R1	R2	cont
    3 008 OP_JMPIF	R2	010
    3 009 OP_JMP	014
    4 010 OP_LOADSELF	R1
    4 011 OP_LOADI	R2	1
    4 012 OP_SEND	R1	:p	1
    4 013 OP_JMP	024
    5 014 OP_GETCONST	R2	:RuntimeError
    5 015 OP_RESCUE	R1	R2	cont
    5 016 OP_JMPIF	R2	018
    5 017 OP_JMP	022
    6 018 OP_LOADSELF	R1
    6 019 OP_LOADI	R2	2
    6 020 OP_SEND	R1	:p	1
    6 021 OP_JMP	024
    6 022 OP_RAISE	R1
    6 023 OP_POPERR	1
    6 024 OP_STOP
```

sprintf: Should use mrb_int for any object

o=Object.new
def o.to_int
  1
end
p "%*d" % [o, 1]

The change may reduce flexibility, but I believe no one wants
that level of flexibility here.

MiniRake::Task#timestamp should use Time.now only

Some update for `String#each_line`

Fix out-of-bound access

Get rid of out-of-bound access when single % at the end.

The argv for `mrb_yield_with_class()` should not be on mruby stack.
Note the result from `mrb_get_args("*")` is on the stack.

The old OP_RESCUE took one operand A, which specifies a class
to match with the exception. The new OP_RESCUE takes tree operands:

A: the register to hold exception
B: the matching exception; the match result will be stored here.
C: the continuation; if C is zero, the exception will be stored to R(A)
   otherwise, the value from R(A) is used as a exception.

Thus,
```ruby
begin
  raise "a"
rescue TypeError
  p 1
rescue RuntimeError
  p 2
end
```
will be compiled as
```
irep 0x557a06667aa0 nregs=4 nlocals=1 pools=1 syms=4 reps=0
file: /tmp/e.rb
2 000 OP_ONERR  005
2 001 OP_LOADSELF R1
2 002 OP_STRING R2  L(0)  ; "a"
2 003 OP_SEND R1  :raise  1
2 004 OP_JMP  022
3 005 OP_GETCONST R2  :TypeError
3 006 OP_RESCUE R1  R2
3 007 OP_JMPIF  R2  009
3 008 OP_JMP  013
4 009 OP_LOADSELF R1
4 010 OP_LOADI  R2  1
4 011 OP_SEND R1  :p  1
4 012 OP_JMP  023
5 013 OP_GETCONST R2 :RuntimeError
5 014 OP_RESCUE R1 R2  cont
5 015 OP_JMPIF  R2  017
5 016 OP_JMP  021
6 017 OP_LOADSELF R1
6 018 OP_LOADI  R2  2
6 019 OP_SEND R1  :p  1
6 020 OP_JMP  023
6 021 OP_RAISE  R1
6 022 OP_POPERR 1
6 023 OP_STOP
```

The new VM can accept old OP_RESCUE. The mruby compatible VM (namely
mruby/c) should be updated to support the new OP_RESCUE behavior.

Also fix #3499.  The issue was solved by #3462.