Ensure span cancellation is done properly with concurrency protection

[kstenerud]

Goal

Using networkRequestCallback_ null value as a flag affecting behaviour is too fragile and has led to multiple bugs. Refactor this section to just re-run the callback for the (max 10 or so) initial spans instead, which makes the concurrency checks far less complex / prone to holes.

As a result:

• isEarlySpansPhase_ and dependent code will never suffer race conditions
• Spans will never end up partially initialized because they ran in between states where the network callback pointed to actual user code.
• The callbacks can never be null

[github-actions]

    FILE SIZE        VM SIZE    
 --------------  -------------- 
  +2.6% +3.98Ki  +2.6% +3.98Ki    String Table
  +1.2%   +1008  +1.2%   +1008    Symbol Table
  +5.2%    +224  +5.2%    +224    __DATA,__const
  +6.2%    +164  +6.2%    +164    __TEXT,__const
  +0.1%    +116  +0.1%    +116    __TEXT,__text
  +1.6%     +85  +1.6%     +85    __TEXT,__cstring
  +1.2%     +44  +1.2%     +44    __TEXT,__unwind_info
  +1.8%     +16  +1.8%     +16    Function Start Addresses
  +0.2%      +8  +0.2%      +8    Binding Info
  +1.1%      +8  +1.1%      +8    Rebase Info
  -0.0%      -1  -0.0%      -1    __TEXT,__objc_methname
  -0.2%     -16  -0.2%     -16    __TEXT,__gcc_except_tab
  -4.3%    -224  -4.4%    -224    [__DATA]
  -1.1%    -392  -1.1%    -392    [__TEXT]
  [ = ]       0 -33.2% -4.99Ki    [__LINKEDIT]
  +1.1% +4.99Ki  [ = ]       0    TOTAL

Generated by 🚫 Danger